An Evaluation of Deep Learning Methods for Small Object Detection

Table 4

The comparative results on subsets of PASCAL VOC 2007.

Approach	Method	VOC_MRA_0.058	VOC_MRA_0.10	VOC_MRA_0.20	VOC_WH20
One stage	YOLOv2 416 [16]	3.02	31.38	42.89	18.52
	YOLOv2 448 [16]	4.47	32.9	60.15	21.96
	YOLOv2 480 [16]	4.26	33.48	60.78	26.67
	YOLOv2 512 [16]	5.42	35.74	61.12	24.63
	YOLOv2 544 [16]	6.97	36.56	63	26.62
	YOLOv2 640 [16]	7.7	37.97	61.29	23.41
	YOLOv2 800 [16]	10.24	37.3	61.91	26.9
	YOLOv2 1024 [16]	10.69	29.93	55.14	28.97
	YOLOv3 320	7.18	34.58	60.36	20.4
	YOLOv3 416	10.2	38.97	62.53	24.12
	YOLOv3 608	11.7	42.65	68.56	28.86
	SSD 300 [16]	1.71	32.76	46.26	16.91
	SSD 512 [16]	2.9	43.46	57.11	19.87
	RetinaNet-ResNet-50-FPN	8.84	41.5	50.2	28.14
	RetinaNet-ResNet-101-FPN	8.95	42.5	51.9	27.46
	RetinaNet-ResNeXT-101-32 × 8d-FPN	10.29	45.4	54.5	30.08
	RetinaNet-ResNeXT-101-64 × 4d-FPN	10.71	45.5	55.1	31.32
Two stage	Fast RCNN-ResNet-50-C4	0.23	13.2	49.9	3.93
	Fast RCNN-ResNet-50-FPN	0.63	13.5	55.6	3.45
	Fast RCNN-ResNet-101-FPN	0.39	15.9	57.6	3.12
	Fast RCNN-ResNeXT-101-32 × 8d-FPN	0.51	14.4	57.9	3.33
	Fast RCNN-ResNeXT-101-64 × 4d-FPN	0.29	14.2	57.3	3.76
	Faster RCNN-ResNet-50-C4	6.98	39.9	48.7	26.04
	Faster RCNN-ResNet-50-FPN	10.74	45.6	56.3	29.79
	Faster RCNN-ResNet-101-FPN	10.63	46.9	57.6	30.57
	Faster RCNN-ResNeXT-101-32 × 8d-FPN	11.64	47.3	57.6	32.12
	Faster RCNN-ResNeXT-101-64 × 4d-FPN	10.54	47.1	56.9	31.64
	Faster RCNN-VGG16 [16]	5.73	35.58	44.14	41.11

This table illustrates how well models adapt to different scales of objects. The values in bold represent the best in one-stage methods, and the ones in italics represent the highest in two-stage methods.