{ "version": "2026-09-09", "description": "Model facts for Model Match: one row per canonical model, keyed \"family|variant\", with the reference accuracy and the metric and dataset it was measured on named rather than assumed, the quantised accuracy per precision with the drop the publisher reported, parameter count, GFLOPs at the input size quoted, weight size per precision (a parameter count multiplied by a byte width, labelled as the model it is), the TensorRT engine / activation / workspace / IO memory split where one is published, and the licence. It also carries the canonical model vocabulary — the alias map that joins the two ids the benchmark database splits one model across, throughput on one and accuracy on the other — and the per-platform, per-precision accuracy figures that the benchmark database drops in its dedupe, each stating which evaluation set it used or recording it as unstated. A vendor model-zoo accuracy table is an external measured benchmark (class C); class B is reserved for EdgeAIStack measurements and is never claimed here. A figure no publisher states is null and the engine reports UNKNOWN rather than guessing; a source without a resolvable url is dropped, and a model left with no sourced fact is dropped and listed in gaps.", "last_updated": "2026-09-09", "confidence": "medium", "evidence_class": "C", "schema": { "models.": "{ id, family, variant, label, task, input_size, params_m { value, class, sources[] }, gflops { value, input_size, class, sources[] }, reference_accuracy { metric, dataset, value, map50, input_size, class, sources[] }, quantised_accuracy. { metric, dataset, value, map50, drop_points, method, calibration, class, sources[] }, weights_mb { fp32, fp16, int8, basis, class, derived, sources[] }, memory_parts. { engine_mb, activation_mb, workspace_mb, calibration_table_mb, io_buffers_mb, total_mb, runtime, measured_on, batch, input_resolution, class, sources[] }, licence, notes[], sources[] }", "aliases.": "{ to, note } — the non-canonical spelling on the left, the canonical model on the right", "platform_accuracy[]": "{ model_id, platform, precision, metric, dataset, value, input_resolution, delta_vs_reference, conflicts_with_reference, class, sources[] }", "metrics": "the accuracy metric vocabulary", "precisions": "the precision vocabulary the per-precision blocks are keyed by", "coverage": "model counts and how many carry each fact", "gaps[]": "facts this pass could not confirm from a published document, plus any model dropped for want of a sourced fact" }, "metrics": [ "map50_95", "map50", "top1" ], "precisions": [ "fp32", "fp16", "int8" ], "coverage": { "models": 33, "by_task": { "detection": 33 }, "with_reference_accuracy": 33, "with_quantised_accuracy": 18, "with_params": 31, "with_gflops": 31, "with_weights": 31, "with_memory_split": 5, "with_licence": 0, "aliases": 29, "platform_accuracy_rows": 32 }, "models": { "efficientdet_lite|lite0": { "id": "efficientdet_lite|lite0", "family": "efficientdet_lite", "variant": "lite0", "label": "EfficientDet-lite0", "task": "detection", "input_size": 320, "params_m": null, "gflops": null, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.2641, "map50": null, "input_size": 320, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet README — EfficientDet-Lite FP32 / INT8 mAP", "doc_id": "github.com/google/automl/blob/master/efficientdet/README.md", "url": "https://github.com/google/automl/blob/master/efficientdet/README.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NobuoTsukamoto (independent)", "document": "EfficientDet-Lite TensorFlow Lite benchmark table", "doc_id": "github.com/NobuoTsukamoto/benchmarks/blob/main/tensorflow/lite/efficentdet/efficientdet.md", "url": "https://github.com/NobuoTsukamoto/benchmarks/blob/main/tensorflow/lite/efficentdet/efficientdet.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.2641, "map50": null, "drop_points": null, "method": "tflite_int8_ptq", "calibration": "representative dataset (unspecified size)", "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet README — EfficientDet-Lite FP32 / INT8 mAP", "doc_id": "github.com/google/automl/blob/master/efficientdet/README.md", "url": "https://github.com/google/automl/blob/master/efficientdet/README.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NobuoTsukamoto (independent)", "document": "EfficientDet-Lite TensorFlow Lite benchmark table", "doc_id": "github.com/NobuoTsukamoto/benchmarks/blob/main/tensorflow/lite/efficentdet/efficientdet.md", "url": "https://github.com/NobuoTsukamoto/benchmarks/blob/main/tensorflow/lite/efficentdet/efficientdet.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.261, "map50": null, "drop_points": 0.31, "method": "tflite_int8_ptq", "calibration": "representative dataset (unspecified size)", "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet README — EfficientDet-Lite FP32 / INT8 mAP", "doc_id": "github.com/google/automl/blob/master/efficientdet/README.md", "url": "https://github.com/google/automl/blob/master/efficientdet/README.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NobuoTsukamoto (independent)", "document": "EfficientDet-Lite TensorFlow Lite benchmark table", "doc_id": "github.com/NobuoTsukamoto/benchmarks/blob/main/tensorflow/lite/efficentdet/efficientdet.md", "url": "https://github.com/NobuoTsukamoto/benchmarks/blob/main/tensorflow/lite/efficentdet/efficientdet.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": null, "memory_parts": {}, "licence": null, "notes": [], "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet README — EfficientDet-Lite FP32 / INT8 mAP", "doc_id": "github.com/google/automl/blob/master/efficientdet/README.md", "url": "https://github.com/google/automl/blob/master/efficientdet/README.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NobuoTsukamoto (independent)", "document": "EfficientDet-Lite TensorFlow Lite benchmark table", "doc_id": "github.com/NobuoTsukamoto/benchmarks/blob/main/tensorflow/lite/efficentdet/efficientdet.md", "url": "https://github.com/NobuoTsukamoto/benchmarks/blob/main/tensorflow/lite/efficentdet/efficientdet.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "efficientdet_lite|lite2": { "id": "efficientdet_lite|lite2", "family": "efficientdet_lite", "variant": "lite2", "label": "EfficientDet-lite2", "task": "detection", "input_size": 448, "params_m": null, "gflops": null, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.3506, "map50": null, "input_size": 448, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet README — EfficientDet-Lite FP32 / INT8 mAP", "doc_id": "github.com/google/automl/blob/master/efficientdet/README.md", "url": "https://github.com/google/automl/blob/master/efficientdet/README.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.3506, "map50": null, "drop_points": null, "method": "tflite_int8_ptq", "calibration": "representative dataset", "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet README — EfficientDet-Lite FP32 / INT8 mAP", "doc_id": "github.com/google/automl/blob/master/efficientdet/README.md", "url": "https://github.com/google/automl/blob/master/efficientdet/README.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.3469, "map50": null, "drop_points": 0.37, "method": "tflite_int8_ptq", "calibration": "representative dataset", "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet README — EfficientDet-Lite FP32 / INT8 mAP", "doc_id": "github.com/google/automl/blob/master/efficientdet/README.md", "url": "https://github.com/google/automl/blob/master/efficientdet/README.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": null, "memory_parts": {}, "licence": null, "notes": [], "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet README — EfficientDet-Lite FP32 / INT8 mAP", "doc_id": "github.com/google/automl/blob/master/efficientdet/README.md", "url": "https://github.com/google/automl/blob/master/efficientdet/README.md", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "efficientdet|d0": { "id": "efficientdet|d0", "family": "efficientdet", "variant": "d0", "label": "EfficientDet-d0", "task": "detection", "input_size": null, "params_m": { "value": 3.9, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 2.54, "input_size": null, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.346, "map50": 0.53, "input_size": null, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 15.6, "fp16": 7.8, "int8": 3.9, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "EfficientNet backbone uses Swish activation (equivalent to SiLU) — similar quantization challenges as YOLOv5. However BiFPN layers quantize well due to weighted feature fusion. INT8 PTQ estimated 2-3 point mAP drop. QAT recommended for production INT8." ], "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "efficientdet|d1": { "id": "efficientdet|d1", "family": "efficientdet", "variant": "d1", "label": "EfficientDet-d1", "task": "detection", "input_size": null, "params_m": { "value": 6.6, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 6.1, "input_size": null, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.405, "map50": 0.59, "input_size": null, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 26.4, "fp16": 13.2, "int8": 6.6, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Similar to D0: Swish activation quantization challenge. BiFPN quantizes well. INT8 PTQ estimated 2-3 point drop. mAP50 estimated." ], "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "efficientdet|d2": { "id": "efficientdet|d2", "family": "efficientdet", "variant": "d2", "label": "EfficientDet-d2", "task": "detection", "input_size": null, "params_m": { "value": 8.1, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 11, "input_size": null, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.439, "map50": 0.625, "input_size": null, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 32.4, "fp16": 16.2, "int8": 8.1, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "INT8 PTQ viable. Slightly dominated by YOLO11s on efficiency (47.0 mAP, 21.5 GFLOPs vs 43.9, 11.0 GFLOPs) but D2 has lower absolute compute. mAP50 estimated." ], "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "efficientdet|d3": { "id": "efficientdet|d3", "family": "efficientdet", "variant": "d3", "label": "EfficientDet-d3", "task": "detection", "input_size": null, "params_m": { "value": 12, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 24.9, "input_size": null, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.472, "map50": 0.658, "input_size": null, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 48, "fp16": 24, "int8": 12, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "INT8 PTQ viable with good calibration. Dominated by YOLO11s in both accuracy and compute. mAP50 estimated." ], "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "efficientdet|d4": { "id": "efficientdet|d4", "family": "efficientdet", "variant": "d4", "label": "EfficientDet-d4", "task": "detection", "input_size": null, "params_m": { "value": 20.7, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 55.2, "input_size": null, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.497, "map50": 0.684, "input_size": null, "class": "C", "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 82.8, "fp16": 41.4, "int8": 20.7, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Larger BiFPN with more layers — slightly harder INT8 calibration. PTQ estimated 2-3 point drop. mAP50 estimated." ], "sources": [ { "publisher": "Google AutoML", "document": "EfficientDet repository benchmark table", "doc_id": "github.com/google/automl/tree/master/efficientdet", "url": "https://github.com/google/automl/tree/master/efficientdet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "nanodet|m-320": { "id": "nanodet|m-320", "family": "nanodet", "variant": "m-320", "label": "NanoDet m-320", "task": "detection", "input_size": 320, "params_m": { "value": 0.95, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 0.72, "input_size": 320, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.206, "map50": 0.32, "input_size": 320, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 3.8, "fp16": 1.8, "int8": 0.98, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Designed for mobile INT8 from ground up. ShuffleNetV2 backbone with depthwise separable convolutions quantize exceptionally well. GFL head with distribution learning quantizes cleanly. INT8 PTQ estimated <1 point drop. Sub-1MB INT8 model — viable for microcontroller with SRAM constraints." ], "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "nanodet|plus-m-1.5x-416": { "id": "nanodet|plus-m-1.5x-416", "family": "nanodet", "variant": "plus-m-1.5x-416", "label": "NanoDet plus-m-1.5x-416", "task": "detection", "input_size": 416, "params_m": { "value": 2.44, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 3.15, "input_size": 416, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.341, "map50": 0.5, "input_size": 416, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 9.8, "fp16": 4.7, "int8": 2.3, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "1.5x width multiplier adds headroom for cleaner INT8 quantization. Recommended as the largest NanoDet variant for embedded targets. INT8 PTQ estimated <1.5 point drop." ], "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "nanodet|plus-m-320": { "id": "nanodet|plus-m-320", "family": "nanodet", "variant": "plus-m-320", "label": "NanoDet plus-m-320", "task": "detection", "input_size": 320, "params_m": { "value": 1.17, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 0.9, "input_size": 320, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.27, "map50": 0.405, "input_size": 320, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 4.7, "fp16": 2.3, "int8": 1.2, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "NanoDet-Plus uses PAN neck and improved head — cleaner feature maps improve INT8 calibration vs original NanoDet. INT8 model at 1.2MB is production-ready for embedded NPU. Estimated PTQ mAP drop <1.5 points." ], "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "nanodet|plus-m-416": { "id": "nanodet|plus-m-416", "family": "nanodet", "variant": "plus-m-416", "label": "NanoDet plus-m-416", "task": "detection", "input_size": 416, "params_m": { "value": 1.17, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 1.52, "input_size": 416, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.304, "map50": 0.452, "input_size": 416, "class": "C", "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 4.7, "fp16": 2.3, "int8": 1.2, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Same model as plus_m_320 at higher input resolution. INT8 viable. Slightly more activation memory at 416px but still sub-4MB peak [est]." ], "sources": [ { "publisher": "RangiLyu (NanoDet authors)", "document": "NanoDet repository — model table", "doc_id": "github.com/RangiLyu/nanodet", "url": "https://github.com/RangiLyu/nanodet", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "rtdetr_v2|l": { "id": "rtdetr_v2|l", "family": "rtdetr_v2", "variant": "l", "label": "RT-DETRv2l", "task": "detection", "input_size": null, "params_m": { "value": 42, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 136, "input_size": null, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.534, "map50": 0.716, "input_size": null, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 168, "fp16": 84, "int8": 42, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "ResNet-50 backbone with transformer decoder. Transformer layers cannot be safely quantized to INT8 without QAT. FP16 is the practical minimum. Jetson AGX Orin FP16 target." ], "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "rtdetr_v2|m": { "id": "rtdetr_v2|m", "family": "rtdetr_v2", "variant": "m", "label": "RT-DETRv2m", "task": "detection", "input_size": null, "params_m": { "value": 36, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 100, "input_size": null, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.519, "map50": 0.699, "input_size": null, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 144, "fp16": 72, "int8": 36, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Same transformer precision sensitivity as RT-DETRv2-S. Mixed-precision recommended. Not suitable for edge INT8 without significant tuning." ], "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "rtdetr_v2|s": { "id": "rtdetr_v2|s", "family": "rtdetr_v2", "variant": "s", "label": "RT-DETRv2s", "task": "detection", "input_size": null, "params_m": { "value": 20, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 60, "input_size": null, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.481, "map50": 0.651, "input_size": null, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 80, "fp16": 40, "int8": 20, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Transformer attention layers require high precision for Q/K/V dot products. Multi-head attention with softmax is highly sensitive to INT8 quantization — normalization across 8-bit values loses critical relative magnitude. Recommended: FP16 for attention layers with mixed-precision INT8 for CNN backbone. PTQ typically drops 3-6 points mAP. QAT improves but requires significant engineering effort." ], "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "rtdetr_v2|x": { "id": "rtdetr_v2|x", "family": "rtdetr_v2", "variant": "x", "label": "RT-DETRv2x", "task": "detection", "input_size": null, "params_m": { "value": 76, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 259, "input_size": null, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.543, "map50": 0.728, "input_size": null, "class": "C", "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 304, "fp16": 152, "int8": 76, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Largest RT-DETR variant. Transformer quantization requires mixed-precision or QAT. Not suitable for edge without Jetson AGX Orin or higher." ], "sources": [ { "publisher": "Lyuwenyu (RT-DETR authors)", "document": "RT-DETR repository — model table", "doc_id": "github.com/lyuwenyu/RT-DETR", "url": "https://github.com/lyuwenyu/RT-DETR", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "ssd_mobilenet|v1": { "id": "ssd_mobilenet|v1", "family": "ssd_mobilenet", "variant": "v1", "label": "SSD MobileNet v1", "task": "detection", "input_size": 300, "params_m": { "value": 6.8, "class": "C", "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 1.14, "input_size": 300, "class": "C", "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.215, "map50": 0.33, "input_size": 300, "class": "C", "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 27.2, "fp16": 13.6, "int8": 7, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "ReLU6 activation (instead of SiLU) quantizes cleanly to INT8. MobileNetV1 depthwise separable convolutions have bounded activation ranges — excellent INT8 PTQ behavior. Designed for TFLite INT8 from inception. PTQ drop typically <0.5 mAP. Reference model for INT8 embedded deployment." ], "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "ssd_mobilenet|v2": { "id": "ssd_mobilenet|v2", "family": "ssd_mobilenet", "variant": "v2", "label": "SSD MobileNet v2", "task": "detection", "input_size": 300, "params_m": { "value": 4.3, "class": "C", "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 0.8, "input_size": 300, "class": "C", "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.221, "map50": null, "input_size": 300, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.221, "map50": null, "drop_points": null, "method": "tflite_int8_ptq", "calibration": "representative dataset from training set", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 17.2, "fp16": 8.6, "int8": 6.6, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "MobileNetV2 inverted residuals with ReLU6 — excellent INT8 quantization. Bottleneck design constrains activation ranges. INT8 PTQ drop estimated <0.5 mAP. Industry standard for TFLite INT8 embedded targets. Recommended for Coral Edge TPU, Arm Ethos-U, NXP i.MX." ], "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "ssdlite_mobilenet|v3-large": { "id": "ssdlite_mobilenet|v3-large", "family": "ssdlite_mobilenet", "variant": "v3-large", "label": "SSDLite MobileNet v3-large", "task": "detection", "input_size": null, "params_m": { "value": 5.1, "class": "C", "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 1, "input_size": null, "class": "C", "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.299, "map50": 0.44, "input_size": null, "class": "C", "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": {}, "weights_mb": { "fp32": 20.4, "fp16": 10.2, "int8": 5.1, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "MobileNetV3 uses hard-swish activation which has bounded range [0, 1] — cleaner INT8 quantization than SiLU/Swish. Lightweight SSDLite detection head (depthwise separable) also quantizes well. INT8 PTQ drop estimated <1 point. TorchVision model, ONNX/TFLite export supported." ], "sources": [ { "publisher": "Google Coral", "document": "Model zoo — object detection", "doc_id": "coral.ai/models/object-detection", "url": "https://coral.ai/models/object-detection/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolo11|l": { "id": "yolo11|l", "family": "yolo11", "variant": "l", "label": "YOLO11l", "task": "detection", "input_size": 640, "params_m": { "value": 25.3, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 86.9, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.534, "map50": 0.692, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.534, "map50": 0.692, "drop_points": null, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.534, "map50": 0.692, "drop_points": 0, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.511, "map50": 0.663, "drop_points": 2.3, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 101.2, "fp16": 50.6, "int8": 25.3, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": { "fp32": { "engine_mb": 101.5, "activation_mb": 136, "workspace_mb": 64, "calibration_table_mb": null, "io_buffers_mb": 4.7, "total_mb": 306, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "D", "sources": [ { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "engine_mb": 47.7, "activation_mb": 70, "workspace_mb": 64, "calibration_table_mb": null, "io_buffers_mb": 2.9, "total_mb": 186, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "C", "sources": [ { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "engine_mb": 33.5, "activation_mb": 70, "workspace_mb": 64, "calibration_table_mb": 3.5, "io_buffers_mb": 2.9, "total_mb": 174, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" }, { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "licence": null, "notes": [ "INT8 PTQ viable. Fewer parameters than YOLOv8l (25.3M vs 43.7M) for comparable accuracy. INT8 estimated 2-3 point drop. Suitable for Jetson AGX Orin FP16 at ~12-15ms [est]." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" } ] }, "yolo11|m": { "id": "yolo11|m", "family": "yolo11", "variant": "m", "label": "YOLO11m", "task": "detection", "input_size": 640, "params_m": { "value": 20.1, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 68, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.515, "map50": 0.672, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.515, "map50": 0.672, "drop_points": null, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.515, "map50": 0.672, "drop_points": 0, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.489, "map50": 0.639, "drop_points": 2.6, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 80.4, "fp16": 40.2, "int8": 20.1, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": { "fp32": { "engine_mb": 80.6, "activation_mb": 110, "workspace_mb": 64, "calibration_table_mb": null, "io_buffers_mb": 4.7, "total_mb": 260, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "D", "sources": [ { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "engine_mb": 38.2, "activation_mb": 58, "workspace_mb": 64, "calibration_table_mb": null, "io_buffers_mb": 2.9, "total_mb": 164, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "C", "sources": [ { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "engine_mb": 27.3, "activation_mb": 58, "workspace_mb": 64, "calibration_table_mb": 3, "io_buffers_mb": 2.9, "total_mb": 156, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" } ] } }, "licence": null, "notes": [ "22% fewer parameters than YOLOv8m for higher accuracy — cleaner INT8 quantization. PTQ estimated 1.5-2.5 point drop. Recommended for Jetson AGX Orin at FP16 (<10ms)." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" } ] }, "yolo11|n": { "id": "yolo11|n", "family": "yolo11", "variant": "n", "label": "YOLO11n", "task": "detection", "input_size": 640, "params_m": { "value": 2.6, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 6.5, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.395, "map50": 0.561, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Science and Information Organization (IJACSA)", "document": "Multi-class object detection using quantized YOLOv11 for real-time inference", "doc_id": "thesai.org/Downloads/Volume16No12/Paper_83-Multi_Class_Object_Detection_Using_Quantized_YOLOv11_for_Real_Time_", "url": "https://thesai.org/Downloads/Volume16No12/Paper_83-Multi_Class_Object_Detection_Using_Quantized_YOLOv11_for_Real_Time_Inference.pdf", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.395, "map50": 0.561, "drop_points": null, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Science and Information Organization (IJACSA)", "document": "Multi-class object detection using quantized YOLOv11 for real-time inference", "doc_id": "thesai.org/Downloads/Volume16No12/Paper_83-Multi_Class_Object_Detection_Using_Quantized_YOLOv11_for_Real_Time_", "url": "https://thesai.org/Downloads/Volume16No12/Paper_83-Multi_Class_Object_Detection_Using_Quantized_YOLOv11_for_Real_Time_Inference.pdf", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.395, "map50": 0.561, "drop_points": 0, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Science and Information Organization (IJACSA)", "document": "Multi-class object detection using quantized YOLOv11 for real-time inference", "doc_id": "thesai.org/Downloads/Volume16No12/Paper_83-Multi_Class_Object_Detection_Using_Quantized_YOLOv11_for_Real_Time_", "url": "https://thesai.org/Downloads/Volume16No12/Paper_83-Multi_Class_Object_Detection_Using_Quantized_YOLOv11_for_Real_Time_Inference.pdf", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.362, "map50": 0.523, "drop_points": 3.3, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Science and Information Organization (IJACSA)", "document": "Multi-class object detection using quantized YOLOv11 for real-time inference", "doc_id": "thesai.org/Downloads/Volume16No12/Paper_83-Multi_Class_Object_Detection_Using_Quantized_YOLOv11_for_Real_Time_", "url": "https://thesai.org/Downloads/Volume16No12/Paper_83-Multi_Class_Object_Detection_Using_Quantized_YOLOv11_for_Real_Time_Inference.pdf", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 10.4, "fp16": 5.2, "int8": 2.6, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": { "fp32": { "engine_mb": 10.6, "activation_mb": 48, "workspace_mb": 64, "calibration_table_mb": null, "io_buffers_mb": 4.7, "total_mb": 128, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "D", "sources": [ { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" } ] }, "fp16": { "engine_mb": 5.3, "activation_mb": 26, "workspace_mb": 64, "calibration_table_mb": null, "io_buffers_mb": 2.9, "total_mb": 98, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" }, { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "engine_mb": 3.5, "activation_mb": 26, "workspace_mb": 64, "calibration_table_mb": 2, "io_buffers_mb": 2.9, "total_mb": 98, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" }, { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "licence": null, "notes": [ "Best nano-class quantization target. C3k2 and C2PSA modules quantize cleanly. INT8 PTQ estimated 1-2 point drop. Extremely low activation memory makes it ideal for INT8 on low-power MCU-class accelerators. T4 latency 1.5ms — achievable sub-5ms INT8 on Jetson Orin Nano." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Science and Information Organization (IJACSA)", "document": "Multi-class object detection using quantized YOLOv11 for real-time inference", "doc_id": "thesai.org/Downloads/Volume16No12/Paper_83-Multi_Class_Object_Detection_Using_Quantized_YOLOv11_for_Real_Time_", "url": "https://thesai.org/Downloads/Volume16No12/Paper_83-Multi_Class_Object_Detection_Using_Quantized_YOLOv11_for_Real_Time_Inference.pdf", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" }, { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolo11|s": { "id": "yolo11|s", "family": "yolo11", "variant": "s", "label": "YOLO11s", "task": "detection", "input_size": 640, "params_m": { "value": 9.4, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 21.5, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.47, "map50": 0.625, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.47, "map50": 0.625, "drop_points": null, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.47, "map50": 0.625, "drop_points": 0, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.44, "map50": 0.588, "drop_points": 3, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 37.6, "fp16": 18.8, "int8": 9.4, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": { "fp32": { "engine_mb": 37.8, "activation_mb": 72, "workspace_mb": 64, "calibration_table_mb": null, "io_buffers_mb": 4.7, "total_mb": 178, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "D", "sources": [ { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" } ] }, "fp16": { "engine_mb": 18.4, "activation_mb": 38, "workspace_mb": 64, "calibration_table_mb": null, "io_buffers_mb": 2.9, "total_mb": 124, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "C", "sources": [ { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "engine_mb": 14.1, "activation_mb": 38, "workspace_mb": 64, "calibration_table_mb": 2.5, "io_buffers_mb": 2.9, "total_mb": 122, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "C", "sources": [ { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" } ] } }, "licence": null, "notes": [ "Excellent PTQ candidate. 21.5 GFLOPs at 47.0 mAP is best small-tier efficiency. INT8 estimated 1.5-2 point drop. Recommended for Jetson Orin Nano INT8 at <5ms target." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" }, { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolo11|x": { "id": "yolo11|x", "family": "yolo11", "variant": "x", "label": "YOLO11x", "task": "detection", "input_size": 640, "params_m": { "value": 56.9, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 194.9, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.547, "map50": 0.71, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.547, "map50": 0.71, "drop_points": null, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.547, "map50": 0.71, "drop_points": 0, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.526, "map50": 0.684, "drop_points": 2.1, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 227.6, "fp16": 113.8, "int8": 56.9, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": { "fp32": { "engine_mb": 228, "activation_mb": 226, "workspace_mb": 128, "calibration_table_mb": null, "io_buffers_mb": 4.7, "total_mb": 588, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "D", "sources": [ { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "engine_mb": 104.2, "activation_mb": 116, "workspace_mb": 128, "calibration_table_mb": null, "io_buffers_mb": 2.9, "total_mb": 352, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "C", "sources": [ { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "engine_mb": 76.5, "activation_mb": 116, "workspace_mb": 128, "calibration_table_mb": 5, "io_buffers_mb": 2.9, "total_mb": 328, "runtime": "tensorrt", "measured_on": "jetson_agx_orin_32gb", "batch": 1, "input_resolution": "640x640", "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" }, { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "licence": null, "notes": [ "Best accuracy in YOLO family. INT8 PTQ estimated 2-4 point drop. Edge server or Jetson AGX Orin only. FP16 recommended for production." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLO11 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolo11", "url": "https://docs.ultralytics.com/models/yolo11/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA (TensorRT samples)", "document": "trtexec profiling output — community benchmark reports", "doc_id": "github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "url": "https://github.com/NVIDIA/TensorRT/tree/main/samples/trtexec", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Qengineering", "document": "Deep learning with Jetson — YOLO benchmarks and memory analysis", "doc_id": "qengineering.eu/deep-learning-with-jetson-nano.html", "url": "https://qengineering.eu/deep-learning-with-jetson-nano.html", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "DeepStream SDK developer guide — memory profiling and allocation reference", "doc_id": "docs.nvidia.com/metropolis/deepstream/dev-guide", "url": "https://docs.nvidia.com/metropolis/deepstream/dev-guide/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "A" } ] }, "yolov5|l": { "id": "yolov5|l", "family": "yolov5", "variant": "l", "label": "YOLOv5l", "task": "detection", "input_size": 640, "params_m": { "value": 53.2, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 135, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.49, "map50": 0.673, "input_size": 640, "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.49, "map50": 0.673, "drop_points": null, "method": "tensorrt_dla_int8_qat", "calibration": "coco_train subset", "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.488, "map50": 0.671, "drop_points": 0.2, "method": "tensorrt_dla_int8_qat", "calibration": "coco_train subset", "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 212.8, "fp16": 106.4, "int8": 53.2, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Large model SiLU+C3 combination creates significant INT8 quantization error. QAT required for <1 point mAP drop. Production deployments should prefer FP16 or use QAT pipeline." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolov5|m": { "id": "yolov5|m", "family": "yolov5", "variant": "m", "label": "YOLOv5m", "task": "detection", "input_size": 640, "params_m": { "value": 25.1, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 64.2, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.454, "map50": 0.641, "input_size": 640, "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.454, "map50": 0.641, "drop_points": null, "method": "tensorrt_ptq_requires_partial_or_qat", "calibration": null, "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.452, "map50": 0.639, "drop_points": 0.2, "method": "tensorrt_ptq_requires_partial_or_qat", "calibration": null, "class": "C", "sources": [ { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 100.4, "fp16": 50.2, "int8": 25.1, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Medium models have wider activation ranges than nano/small — INT8 PTQ degradation can reach 3-6 points. QAT needed for production INT8." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolov5|n": { "id": "yolov5|n", "family": "yolov5", "variant": "n", "label": "YOLOv5n", "task": "detection", "input_size": 640, "params_m": { "value": 2.6, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 7.7, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.28, "map50": 0.457, "input_size": 640, "class": "C", "sources": [ { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.28, "map50": 0.457, "drop_points": null, "method": "tensorrt_ptq_requires_partial_or_qat", "calibration": null, "class": "C", "sources": [ { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.279, "map50": 0.455, "drop_points": 0.1, "method": "tensorrt_ptq_requires_partial_or_qat", "calibration": null, "class": "C", "sources": [ { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 10.4, "fp16": 5.2, "int8": 2.6, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "SiLU activation causes outlier activations that degrade INT8 PTQ quality in TensorRT. Typical PTQ mAP drop 2-4 points. QAT recovers ~1-2 points. C3 bottleneck residuals also create wide activation ranges. Anchor-based head with 3 scale outputs adds quantization complexity." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolov5|s": { "id": "yolov5|s", "family": "yolov5", "variant": "s", "label": "YOLOv5s", "task": "detection", "input_size": 640, "params_m": { "value": 9.1, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 24, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.367, "map50": 0.558, "input_size": 640, "class": "C", "sources": [ { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA (TensorRT issue tracker)", "document": "TensorRT issue 1114 — YOLOv5s INT8 accuracy collapse analysis", "doc_id": "github.com/NVIDIA/TensorRT/issues/1114", "url": "https://github.com/NVIDIA/TensorRT/issues/1114", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.367, "map50": 0.558, "drop_points": null, "method": "tensorrt_ptq_partial_and_qat", "calibration": "coco_train subset", "class": "C", "sources": [ { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA (TensorRT issue tracker)", "document": "TensorRT issue 1114 — YOLOv5s INT8 accuracy collapse analysis", "doc_id": "github.com/NVIDIA/TensorRT/issues/1114", "url": "https://github.com/NVIDIA/TensorRT/issues/1114", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.365, "map50": 0.556, "drop_points": 0.2, "method": "tensorrt_ptq_partial_and_qat", "calibration": "coco_train subset", "class": "C", "sources": [ { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA (TensorRT issue tracker)", "document": "TensorRT issue 1114 — YOLOv5s INT8 accuracy collapse analysis", "doc_id": "github.com/NVIDIA/TensorRT/issues/1114", "url": "https://github.com/NVIDIA/TensorRT/issues/1114", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 36.4, "fp16": 18.2, "int8": 9.1, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Same SiLU quantization challenge as YOLOv5n. INT8 PTQ typically drops 2-5 mAP points on COCO. QAT with NVIDIA TensorRT toolkit can recover most accuracy. Recommended: use FP16 if latency headroom allows." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA (TensorRT issue tracker)", "document": "TensorRT issue 1114 — YOLOv5s INT8 accuracy collapse analysis", "doc_id": "github.com/NVIDIA/TensorRT/issues/1114", "url": "https://github.com/NVIDIA/TensorRT/issues/1114", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "NVIDIA", "document": "Deploying YOLOv5 on Jetson Orin with cuDLA quantization-aware training", "doc_id": "developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inf", "url": "https://developer.nvidia.com/blog/deploying-yolov5-on-nvidia-jetson-orin-with-cudla-quantization-aware-training-to-inference/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolov5|x": { "id": "yolov5|x", "family": "yolov5", "variant": "x", "label": "YOLOv5x", "task": "detection", "input_size": 640, "params_m": { "value": 97.2, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 246.4, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.507, "map50": 0.689, "input_size": 640, "class": "C", "sources": [ { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.507, "map50": 0.689, "drop_points": null, "method": "tensorrt_ptq_requires_partial_or_qat", "calibration": null, "class": "C", "sources": [ { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.505, "map50": 0.687, "drop_points": 0.2, "method": "tensorrt_ptq_requires_partial_or_qat", "calibration": null, "class": "C", "sources": [ { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 388.8, "fp16": 194.4, "int8": 97.2, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Heaviest YOLOv5 variant. INT8 PTQ drops ~5 points mAP on COCO. Not recommended for edge INT8 without QAT. Activation memory at 640px input is ~200-300MB — challenging for edge devices." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLOv5 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov5", "url": "https://docs.ultralytics.com/models/yolov5/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "maggiez0138 (independent)", "document": "yolov5_quant_sample — YOLOv5s quantisation mAP table", "doc_id": "github.com/maggiez0138/yolov5_quant_sample", "url": "https://github.com/maggiez0138/yolov5_quant_sample", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolov8|l": { "id": "yolov8|l", "family": "yolov8", "variant": "l", "label": "YOLOv8l", "task": "detection", "input_size": 640, "params_m": { "value": 43.7, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 165.2, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.529, "map50": 0.657, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.529, "map50": 0.657, "drop_points": null, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.529, "map50": 0.657, "drop_points": 0, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.497, "map50": 0.615, "drop_points": 3.2, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 174.8, "fp16": 87.4, "int8": 43.7, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "INT8 PTQ viable with careful calibration. FP16 preferred for production. Dominated by YOLO11l which achieves 53.4 mAP at 86.9 GFLOPs — half the compute." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolov8|m": { "id": "yolov8|m", "family": "yolov8", "variant": "m", "label": "YOLOv8m", "task": "detection", "input_size": 640, "params_m": { "value": 25.9, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 78.9, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.502, "map50": 0.632, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.502, "map50": 0.632, "drop_points": null, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.502, "map50": 0.632, "drop_points": 0, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.467, "map50": 0.583, "drop_points": 3.5, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 103.6, "fp16": 51.8, "int8": 25.9, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Reasonable INT8 PTQ behavior. Slightly dominated by YOLO11m which achieves 51.5 mAP at 68 GFLOPs vs 78.9. For INT8 deployment prefer YOLO11m." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolov8|n": { "id": "yolov8|n", "family": "yolov8", "variant": "n", "label": "YOLOv8n", "task": "detection", "input_size": 640, "params_m": { "value": 3.2, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 8.7, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.373, "map50": 0.52, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.373, "map50": 0.52, "drop_points": null, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.373, "map50": 0.52, "drop_points": 0, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.331, "map50": 0.472, "drop_points": 4.2, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 12.8, "fp16": 6.4, "int8": 3.2, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Anchor-free C2f architecture with SiLU. INT8 PTQ via TensorRT: ~0.04 mAP50-95 drop (37.3 → 33.0 [est]). FP16 maintains near-lossless accuracy. Best quantization behavior in YOLOv8 family due to fewer parameters — less activation memory, tighter ranges." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolov8|s": { "id": "yolov8|s", "family": "yolov8", "variant": "s", "label": "YOLOv8s", "task": "detection", "input_size": 640, "params_m": { "value": 11.2, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 28.6, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.449, "map50": 0.579, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "MDPI Computers", "document": "YOLOv8 on Jetson Orin NX — peer-reviewed benchmark", "doc_id": "mdpi.com/2073-431X/15/2/74", "url": "https://www.mdpi.com/2073-431X/15/2/74", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.449, "map50": 0.579, "drop_points": null, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "MDPI Computers", "document": "YOLOv8 on Jetson Orin NX — peer-reviewed benchmark", "doc_id": "mdpi.com/2073-431X/15/2/74", "url": "https://www.mdpi.com/2073-431X/15/2/74", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.449, "map50": 0.579, "drop_points": 0, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "MDPI Computers", "document": "YOLOv8 on Jetson Orin NX — peer-reviewed benchmark", "doc_id": "mdpi.com/2073-431X/15/2/74", "url": "https://www.mdpi.com/2073-431X/15/2/74", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.41, "map50": 0.525, "drop_points": 3.9, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "MDPI Computers", "document": "YOLOv8 on Jetson Orin NX — peer-reviewed benchmark", "doc_id": "mdpi.com/2073-431X/15/2/74", "url": "https://www.mdpi.com/2073-431X/15/2/74", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 44.8, "fp16": 22.4, "int8": 11.2, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "Good INT8 PTQ behavior. FP16 maintains full mAP. INT8 PTQ estimated 1.5-2.5 point drop. Suitable for Jetson Orin Nano INT8 at ~3-5ms. Recommended PTQ calibration with ≥500 representative images." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "MDPI Computers", "document": "YOLOv8 on Jetson Orin NX — peer-reviewed benchmark", "doc_id": "mdpi.com/2073-431X/15/2/74", "url": "https://www.mdpi.com/2073-431X/15/2/74", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "yolov8|x": { "id": "yolov8|x", "family": "yolov8", "variant": "x", "label": "YOLOv8x", "task": "detection", "input_size": 640, "params_m": { "value": 68.2, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "gflops": { "value": 257.8, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "reference_accuracy": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.539, "map50": 0.669, "input_size": 640, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "quantised_accuracy": { "fp32": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.539, "map50": 0.669, "drop_points": null, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "fp16": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.539, "map50": 0.669, "drop_points": 0, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "int8": { "metric": "map50_95", "dataset": "coco_val2017", "value": 0.509, "map50": 0.63, "drop_points": 3, "method": "tensorrt_ptq", "calibration": "coco_train subset 1000 images", "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "weights_mb": { "fp32": 272.8, "fp16": 136.4, "int8": 68.2, "basis": "parameter count x bytes per parameter (4 / 2 / 1). Weight-only: activations and the TensorRT engine overhead are not in it.", "class": "D", "derived": true, "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, "memory_parts": {}, "licence": null, "notes": [ "INT8 PTQ can be used but expect 2-4 point mAP drop. FP16 is standard for edge server class. At 257.8 GFLOPs, not viable for constrained edge without INT8." ], "sources": [ { "publisher": "Ultralytics", "document": "YOLOv8 model page — performance table", "doc_id": "docs.ultralytics.com/models/yolov8", "url": "https://docs.ultralytics.com/models/yolov8/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" }, { "publisher": "Ultralytics", "document": "TensorRT integration — FP32 / FP16 / INT8 accuracy table", "doc_id": "docs.ultralytics.com/integrations/tensorrt", "url": "https://docs.ultralytics.com/integrations/tensorrt/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } }, "aliases": { "yolov8_seg|n-seg": { "to": "yolov8_seg|n", "note": "Task suffix: the same YOLOv8n-seg checkpoint as yolov8_seg|n; the throughput rows are on one id and the accuracy row on the other." }, "yolov8_seg|s-seg": { "to": "yolov8_seg|s", "note": "Task suffix: the same YOLOv8s-seg checkpoint as yolov8_seg|s." }, "yolov8_seg|m-seg": { "to": "yolov8_seg|m", "note": "Task suffix: the same YOLOv8m-seg checkpoint as yolov8_seg|m." }, "yolo11_seg|n-seg": { "to": "yolo11_seg|n", "note": "Task suffix: the same YOLO11n-seg checkpoint as yolo11_seg|n." }, "yolo11_seg|m-seg": { "to": "yolo11_seg|m", "note": "Task suffix: the same YOLO11m-seg checkpoint as yolo11_seg|m." }, "yolov8_pose|n-pose": { "to": "yolov8_pose|n", "note": "Task suffix: the same YOLOv8n-pose checkpoint as yolov8_pose|n." }, "yolov8_pose|s-pose": { "to": "yolov8_pose|s", "note": "Task suffix: the same YOLOv8s-pose checkpoint as yolov8_pose|s." }, "yolov8_pose|m-pose": { "to": "yolov8_pose|m", "note": "Task suffix: the same YOLOv8m-pose checkpoint as yolov8_pose|m." }, "yolo11_pose|n-pose": { "to": "yolo11_pose|n", "note": "Task suffix: the same YOLO11n-pose checkpoint as yolo11_pose|n." }, "yolo11_pose|m-pose": { "to": "yolo11_pose|m", "note": "Task suffix: the same YOLO11m-pose checkpoint as yolo11_pose|m." }, "rtdetr|rtdetr-l": { "to": "rtdetr|l", "note": "Variant spelling: the family name repeated inside the variant." }, "rtdetr|rtdetr-x": { "to": "rtdetr|x", "note": "Variant spelling: the family name repeated inside the variant." }, "efficientdet|lite0": { "to": "efficientdet_lite|lite0", "note": "EfficientDet-Lite is a different architecture from EfficientDet-D; the Lite rows are collected under efficientdet_lite." }, "efficientdet|lite1": { "to": "efficientdet_lite|lite1", "note": "EfficientDet-Lite rows are collected under efficientdet_lite." }, "efficientdet|lite2": { "to": "efficientdet_lite|lite2", "note": "EfficientDet-Lite rows are collected under efficientdet_lite." }, "efficientdet|lite3": { "to": "efficientdet_lite|lite3", "note": "EfficientDet-Lite rows are collected under efficientdet_lite." }, "efficientnetedgetpu|s": { "to": "efficientnet_edgetpu|s", "note": "Family spelling with and without the separator." }, "efficientnetedgetpu|m": { "to": "efficientnet_edgetpu|m", "note": "Family spelling with and without the separator." }, "efficientnetedgetpu|l": { "to": "efficientnet_edgetpu|l", "note": "Family spelling with and without the separator." }, "efficientnet|edgetpu-s": { "to": "efficientnet_edgetpu|s", "note": "EfficientNet-EdgeTPU spelled as a variant of efficientnet." }, "efficientnet|edgetpu-m": { "to": "efficientnet_edgetpu|m", "note": "EfficientNet-EdgeTPU spelled as a variant of efficientnet." }, "efficientnet|edgetpu-l": { "to": "efficientnet_edgetpu|l", "note": "EfficientNet-EdgeTPU spelled as a variant of efficientnet." }, "mobilenet_ssd|v1": { "to": "ssd_mobilenet|v1", "note": "Three spellings of SSD MobileNet v1 exist in the database; ssd_mobilenet is canonical." }, "mobilenetssd|mobilenet-v1-ssd": { "to": "ssd_mobilenet|v1", "note": "Three spellings of SSD MobileNet v1 exist in the database; ssd_mobilenet is canonical." }, "mobilenet|v1-ssd": { "to": "ssd_mobilenet|v1", "note": "SSD MobileNet v1 spelled as a variant of mobilenet." }, "mobilenet|ssd-mobilenet-v1": { "to": "ssd_mobilenet|v1", "note": "SSD MobileNet v1 spelled as a variant of mobilenet." }, "mobilenetssd|mobilenet-v2-ssd": { "to": "ssd_mobilenet|v2", "note": "SSD MobileNet v2 spelled under the mobilenetssd family." }, "mobilenet|ssd-mobilenet-v2": { "to": "ssd_mobilenet|v2", "note": "SSD MobileNet v2 spelled as a variant of mobilenet." }, "deeplab|v3": { "to": "deeplabv3|mobilenetv2", "note": "DeepLabV3 MobileNetV2 under two family spellings; the tasks on the two ids disagree, so the merged candidate carries both." } }, "platform_accuracy": [ { "model_id": "yolo11|l", "platform": "jetson_agx_orin", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.6376, "input_resolution": "640x640", "delta_vs_reference": 0.1036, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|l", "platform": "jetson_agx_orin", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.6208, "input_resolution": "640x640", "delta_vs_reference": 0.0868, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|l", "platform": "jetson_orin_nano_super", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.6361, "input_resolution": "640x640", "delta_vs_reference": 0.1021, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|l", "platform": "jetson_orin_nano_super", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.6207, "input_resolution": "640x640", "delta_vs_reference": 0.0867, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|l", "platform": "jetson_orin_nx", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.6363, "input_resolution": "640x640", "delta_vs_reference": 0.1023, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|l", "platform": "jetson_orin_nx", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.6207, "input_resolution": "640x640", "delta_vs_reference": 0.0867, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|m", "platform": "jetson_agx_orin", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.6305, "input_resolution": "640x640", "delta_vs_reference": 0.1155, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|m", "platform": "jetson_agx_orin", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.6291, "input_resolution": "640x640", "delta_vs_reference": 0.1141, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|m", "platform": "jetson_orin_nano_super", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.6309, "input_resolution": "640x640", "delta_vs_reference": 0.1159, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|m", "platform": "jetson_orin_nano_super", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.6291, "input_resolution": "640x640", "delta_vs_reference": 0.1141, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|m", "platform": "jetson_orin_nx", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.6302, "input_resolution": "640x640", "delta_vs_reference": 0.1152, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|m", "platform": "jetson_orin_nx", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.6291, "input_resolution": "640x640", "delta_vs_reference": 0.1141, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|n", "platform": "jetson_agx_orin", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.5063, "input_resolution": "640x640", "delta_vs_reference": 0.1113, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|n", "platform": "jetson_agx_orin", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.4719, "input_resolution": "640x640", "delta_vs_reference": 0.0769, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|n", "platform": "jetson_orin_nano_super", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.5061, "input_resolution": "640x640", "delta_vs_reference": 0.1111, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|n", "platform": "jetson_orin_nano_super", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.4825, "input_resolution": "640x640", "delta_vs_reference": 0.0875, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|n", "platform": "jetson_orin_nx", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.5085, "input_resolution": "640x640", "delta_vs_reference": 0.1135, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|n", "platform": "jetson_orin_nx", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.4786, "input_resolution": "640x640", "delta_vs_reference": 0.0836, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|s", "platform": "jetson_agx_orin", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.5803, "input_resolution": "640x640", "delta_vs_reference": 0.1103, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|s", "platform": "jetson_agx_orin", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.5735, "input_resolution": "640x640", "delta_vs_reference": 0.1035, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|s", "platform": "jetson_orin_nano_super", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.5779, "input_resolution": "640x640", "delta_vs_reference": 0.1079, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|s", "platform": "jetson_orin_nano_super", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.5735, "input_resolution": "640x640", "delta_vs_reference": 0.1035, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|s", "platform": "jetson_orin_nx", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.5796, "input_resolution": "640x640", "delta_vs_reference": 0.1096, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|s", "platform": "jetson_orin_nx", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.5735, "input_resolution": "640x640", "delta_vs_reference": 0.1035, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|x", "platform": "jetson_agx_orin", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.6887, "input_resolution": "640x640", "delta_vs_reference": 0.1417, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|x", "platform": "jetson_agx_orin", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.6574, "input_resolution": "640x640", "delta_vs_reference": 0.1104, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|x", "platform": "jetson_orin_nano_super", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.6892, "input_resolution": "640x640", "delta_vs_reference": 0.1422, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|x", "platform": "jetson_orin_nano_super", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.6612, "input_resolution": "640x640", "delta_vs_reference": 0.1142, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|x", "platform": "jetson_orin_nx", "precision": "fp16", "metric": "map50_95", "dataset": "unstated", "value": 0.6885, "input_resolution": "640x640", "delta_vs_reference": 0.1415, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo11|x", "platform": "jetson_orin_nx", "precision": "int8", "metric": "map50_95", "dataset": "unstated", "value": 0.6592, "input_resolution": "640x640", "delta_vs_reference": 0.1122, "conflicts_with_reference": true, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "NVIDIA Jetson guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/nvidia-jetson", "url": "https://docs.ultralytics.com/guides/nvidia-jetson/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo26|n", "platform": "raspberry_pi_5", "precision": "fp32", "metric": "map50_95", "dataset": "unstated", "value": 0.4805, "input_resolution": "640x640", "delta_vs_reference": null, "conflicts_with_reference": false, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "Raspberry Pi guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/raspberry-pi", "url": "https://docs.ultralytics.com/guides/raspberry-pi/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] }, { "model_id": "yolo26|n", "platform": "raspberry_pi_5", "precision": "fp32", "metric": "map50_95", "dataset": "unstated", "value": 0.4818, "input_resolution": "640x640", "delta_vs_reference": null, "conflicts_with_reference": false, "class": "C", "sources": [ { "publisher": "Ultralytics", "document": "Raspberry Pi guide — YOLO benchmarks", "doc_id": "docs.ultralytics.com/guides/raspberry-pi", "url": "https://docs.ultralytics.com/guides/raspberry-pi/", "retrieved": "2026-09-09", "verified": "2026-09-09", "class": "C" } ] } ], "gaps": [ "12 accuracy rows in the benchmark corpus that the benchmark database drops carry no source with a resolvable url — their only provenance is a sentence describing how they were scaled from another row. They were not transcribed.", "Licence is not carried for any model: no vendor licence statement was transcribed in this pass, so the licence filter has nothing to filter on.", "Classification top-1 accuracy is not carried: the only top-1 values in this repository sit in a research file whose source records have no url, so they are not click-checkable and were not promoted.", "Speech-recognition word error rate and language-model quality scores are not carried: no evaluation figure for the Whisper models or for any registry LLM or VLM exists in this repository.", "GFLOPs and parameters are carried only for the detector families the model-efficiency file covers; the classification, pose, segmentation and transformer tails have neither.", "Input size is null wherever no input file states it (RT-DETRv2 and EfficientDet-D), rather than being inferred from the family.", "Weight size per precision is a parameter count multiplied by a byte width, not a published artefact size: it is a model, and the registry says so on every row.", "The activation / workspace / IO memory split exists for YOLO11 n-x only; every other detector reports a modelled estimate instead.", "The Jetson-guide per-platform accuracy figures do not state which evaluation set they used and sit 0.10-0.13 above the same model card mAP50-95, so their dataset is recorded as unstated and they are never used as a reference accuracy.", "yolov8_seg|n: dropped — no fact in this pass carries a source with a resolvable url", "yolov8_seg|s: dropped — no fact in this pass carries a source with a resolvable url" ] }