Object Detection Based on Swin Deformable Transformer-BiPAFPN-YOLOX

Table 6

Comparison of our networks with state-of-the-art methods on the COCO 2017 test set. Bold numbers indicate the best results, and blue numbers indicate the second best results. Swin DeTr denotes Swin Deformable Transformer.

Methods	Backbone	Year	FPS	AP (%)	AP50	AP75	APS	APM	APL	Param (M)	GFLOPs	Infer time (ms)
YOLOv5 [4]	Modified CSPv5	2021	115.0	36.7	62.4	44.1	20.5	45.7	51.9	7.3	17.1	8.7
YOLOX [2]	Modified CSPv5	2021	102.0	39.6	64.6	47.5	22.7	48.4	54.1	9.0	26.8	9.8
YOLOX [23]	Swin transformer v2	2022	84.6	43.1	64.5	48.6	23.7	49.6	57.9	9.3	34.6	11.8
YOLOX [17]	PVTv2	2022	82.7	42.8	63.9	48.4	23.3	49.0	58.1	9.7	36.8	12.1
YOLOX-S	Swin DeTr (ours)	—	122.0	44.7	67.7	50.3	25.9	50.9	59.6	7.1	16.3	8.2
YOLOv5-M [4]	Modified CSPv5	2021	90.1	44.5	63.1	—	—	—	—	21.4	51.4	11.1
YOLOX [2]	Modified CSPv5	2021	81.3	46.4	65.4	50.6	26.3	51.0	59.9	25.3	73.8	12.3
YOLOv4-CSP [4]	Modified CSP	2020	73.0	47.5	66.2	51.7	28.2	51.2	59.8	26.7	94.4	13.7
YOLOX [23]	Swin transformer v2	2022	70.1	46.6	67.1	51.9	28.6	51.3	59.6	28.1	96.2	14.3
YOLOX-M [17]	PVTv2	2022	66.7	46.1	66.9	50.8	28.9	51.6	60.1	29.5	101.6	15.0
YOLOX-M	Swin DeTr (ours)	—	104.3	48.4	69.3	53.7	28.7	52.4	61.2	21.2	50.6	9.6
Reference [54]	Darknet-53	2021	95.2	44.3	64.6	—	—	—	—	63.0	177.3	10.5
YOLOX [2]	Darknet-53	2021	90.1	47.4	67.3	52.1	27.5	51.5	60.9	63.7	185.3	11.1
YOLOv5-L [4]	Modified CSPv5	2021	73.0	48.2	66.9	—	—	—	—	65.1	188.6	13.7
YOLOX-L [2]	Modified CSPv5	2021	69.0	50.0	68.5	54.5	29.8	54.5	64.4	68.2	195.6	14.5
PP-YOLOv2 [47]	ResNet50-vd-dcn	2021	68.9	49.5	68.2	54.4	30.7	52.9	61.2	68.8	197.6	14.5
YOLOX-L [23]	Swin transformer v2	2022	53.9	48.5	67.6	54.8	29.9	54.8	66.5	74.1	206.1	18.6
YOLOX-L [17]	PVTv2	2022	52.7	48.1	67.8	53.9	29.4	53.3	64.6	74.7	208.4	19.0
Def DETR [8]	ResNeXt-101	2021	50.8	49.0	68.5	53.2	29.7	51.7	62.8	76.4	209.8	19.7
YOLOX-L	Swin DeTr (ours)	—	93.0	51.8	69.6	55.4	31.7	55.8	66.0	63.5	181.7	10.8
YOLOv4 [3]	CSPDarknet-53	2020	64.0	43.5	65.7	47.3	26.7	46.7	53.3	86.5	223.5	15.6
YOLOv5-X [4]	Modified CSPv5	2021	62.5	50.4	68.8	—	—	—	—	87.8	239.0	16.0
PP-YOLOv2 [47]	ResNet101-vd-dcn	2021	59.3	50.3	69.0	55.3	31.6	53.9	62.4	90.7	267.1	16.9
YOLOX-X [2]	Modified CSPv5	2021	57.8	51.2	69.6	55.7	31.2	56.1	66.1	99.1	286.9	17.3
YOLOX-X [23]	Swin transformer v2	2022	40.5	50.6	69.1	56.2	31.1	55.7	67.1	110.7	292.5	24.7
YOLOX-X [17]	PVTv2	2022	38.8	50.1	68.9	54.6	30.4	55.9	65.2	113.4	295.7	25.8
YOLOX-X	Swin DeTr (ours)	—	71.0	52.1	70.4	57.8	31.9	56.9	66.7	85.5	225.4	14.1