Object Detection Based on Swin Deformable Transformer-BiPAFPN-YOLOX

Table 4

Comparison of training results of Swin DeTr-BiPAFPN-YOLOX, Swin Tr-BiPAFPN-YOLOX and others on the COCO 2017 test set. Swin Tr denotes Swin Transformer, Swin DeTr denotes Swin Deformable Transformer.

Method	Epochs	AP	AP50	AP75	APS	APM	APL	Param (M)	GFLOPs	Infer time (ms)	FPS
YOLOv4 [3]	300	43.5	65.7	47.3	26.7	46.7	53.3	86.5	223.5	15.6	64.0
YOLOv5 [4]	300	44.5	63.1	—	—	—	—	21.4	51.4	11.1	90.1
DETR [7]	500	42.0	62.4	44.2	20.5	45.8	61.1	41.4	86.9	35.7	28.0
DarkNet53-PAFPN-YOLOX [2]	300	47.4	67.3	52.1	27.5	51.5	60.9	63.7	185.3	11.1	90.1
Swin Tr-BiPAFPN-YOLOX	350	48.4	67.8	52.6	29.3	52.6	61.8	79.4	211.8	13.7	73.0
Swin DeTr-BiPAFPN-YOLOX	156	49.7	68.6	53.2	31.2	54.9	63.3	44.5	110.7	10.1	98.7