Benchmarks

This page summarizes benchmark runs generated with opendetect-benchmark.

Definitions:

• Pipeline FPS: end-to-end runtime excluding model load (preprocess + inference + postprocess)

• Infer FPS: model forward pass only

• mAP: reported COCO mAP for each model variant

Plots map:

• x: inference FPS (log scale)

• y: COCO mAP

• marker shape: runtime (CPU/CoreML/CUDA/TRT)

• stroke color: model family

• fill color: model variant

• bubble size: parameter count

RF-DETR

Model	Params (M)	FLOPs (G)	COCO mAP
RF-DETR-n	26.877	31.787	48.4
RF-DETR-s	28.524	59.560	53.0
RF-DETR-m	30.101	78.492	54.7
RF-DETR-l	30.350	125.234	56.5

Model	CPU (Pipeline/Infer)	CoreML (Pipeline/Infer)	CUDA (Pipeline/Infer)	TRT FP32 (Pipeline/Infer)	TRT FP16 (Pipeline/Infer)
RF-DETR-n	20.95 / 22.15	39.95 / 44.04	128.48 / 155.59	213.20 / 301.27	348.89 / 665.33
RF-DETR-s	11.48 / 12.01	28.80 / 32.09	80.82 / 97.25	126.40 / 170.44	224.73 / 425.00
RF-DETR-m	8.15 / 8.49	22.19 / 24.53	62.53 / 74.13	93.42 / 121.64	189.81 / 365.26
RF-DETR-l	4.76 / 4.92	14.74 / 16.20	36.66 / 43.24	55.10 / 69.67	126.76 / 246.85

YOLOX

Model	Params (M)	FLOPs (G)	COCO mAP
YOLOX-tiny	5.061	6.413	32.8
YOLOX-s	8.990	26.686	40.5
YOLOX-m	25.338	73.530	46.9
YOLOX-l	54.208	155.293	49.7
YOLOX-x	99.055	281.410	51.1

Model	CPU (Pipeline/Infer)	CoreML (Pipeline/Infer)	CUDA (Pipeline/Infer)	TRT FP32 (Pipeline/Infer)	TRT FP16 (Pipeline/Infer)
YOLOX-tiny	54.97 / 57.05	130.92 / 142.59	197.38 / 206.49	522.87 / 588.58	776.84 / 935.37
YOLOX-s	18.00 / 18.00	72.30 / 78.24	105.81 / 108.35	n/a	n/a
YOLOX-m	8.11 / 8.19	49.12 / 51.73	69.66 / 70.77	n/a	n/a
YOLOX-l	4.94 / 4.97	35.07 / 36.31	46.28 / 46.78	n/a	n/a
YOLOX-x	3.04 / 3.05	22.23 / 22.73	28.69 / 28.89	n/a	n/a

Notes

• YOLOX m/l/x are currently not supported on TensorRT in this project.

• Benchmark numbers depend on export graph, runtime versions, and hardware/software stack.