Object Detection Benchmark

Based on Mqbench and EOD , we provide an object detection benchmark on COCO dataset. We test the two most popular object detection methods, retinanet and yolox on several Backends.

Backend	w_calibration	a_calibration	w_fakequantize	a_fakequantize	wbit	abit	retinanet			yolox
Backend	w_calibration	a_calibration	w_fakequantize	a_fakequantize	wbit	abit	pretrained@float32	calibration@int8	qat@int8	pretrained@float32	calibration@int8	qat@int8
tensorrt	MinMax	EMAMinMax	Fixed	Fixed	8	8	40.7	40.5	40.7	40.5	39.4	39.8
snpe	MinMax	EMAMinMax	Fixed	Fixed	8	8	40.7	39.7	40.2	40.5	38.1	39.8
vitis	MinMaxFloor	ModeMinMaxFloor	Tqt	Tqt	8	8	40.7	39.0	40.1	29.3*	25.3	27.4

Note*: We provide an simplified model of yolox to meet vitis operator support, which is winner solution for the Low Power Computer Vision Challenge 2021 (LPCV2021). The model is yolox-lpcv.