An Evaluation of Deep Learning Methods for Small Object Detection

Table 3

Comparative results on small object dataset.

Method	Backbone	Clock	Faucet	Jar	Mouse	Outlet	Plate	Switch	Tel.	t. box	t. paper	mAP
YOLO 416 [16]	Darknet-19	22.8	30.8	4	52	20.4	13.1	13	6.1	0	35.3	19.39
YOLO 448 [16]		23	36.9	9	52.5	18.4	13.6	17.5	4.2	0	34.3	20.13
YOLO 480 [16]		34.2	37.3	9.1	53.3	21.4	13.6	15.8	9.1	9.1	34.2	23.71
YOLO 512 [16]		23.1	36.6	6.1	59.8	24.6	14.2	15.7	9.1	4.5	32.4	22.61
YOLO 554 [16]		23.4	37.2	9.1	60.1	27.2	13.4	19.9	9.1	4.5	34.5	23.84
YOLO 640 [16]		20.2	36.2	3.2	59.8	27.8	11.7	18.1	8.2	4.5	35.6	22.53
YOLO 800 [16]		27.6	36	2.3	60.2	32.8	13.1	23.3	9.1	9.1	26.7	24.02
YOLO 1024 [16]		21.7	29.3	1.4	58.3	26.4	11.8	17.5	9.1	9.1	15.7	20.03
YOLO 320	Darknet-53	26.22	38.38	4.55	56.46	36.42	13.34	24.8	10.65	4.55	42.96	25.83
YOLO 416		28.47	47.15	10.83	60.49	43.15	15.87	30.73	15.15	2.62	48.3	30.28
YOLO 608		29.98	47.89	10.76	65.88	48.02	18.09	31.22	14.62	17.99	46.56	33.1
YOLO 320	ResNet-50	19.57	25.73	0.67	45.17	14.37	9.38	13.84	9.09	9.09	23.7	17.06
YOLO 416		23.78	36.65	0.4	54.23	18.37	13.75	19.78	9.84	9.42	35.68	22.19
YOLO 608		26.92	40.65	1.77	61.86	29.18	15.04	20.24	10.09	13.29	36.01	25.5
YOLO 320	ResNet-101	20.52	27.9	0.57	44.68	16.98	13.05	13.66	9.66	9.09	24.36	18.05
YOLO 416		25.72	35.6	3.03	55.73	22.4	15.61	17.26	9.32	3.03	38.71	22.64
YOLO 608		28.79	44.59	9.42	62.18	33.34	15.53	23.88	13.24	15.83	39.17	28.6
YOLO 320	ResNet-152	21.64	27.56	3.03	48.06	17.39	11.12	14.51	9.09	4.55	31.88	18.88
YOLO 416		25.7	36.54	0.89	53.81	20.6	14.13	20.21	11.49	0.29	33.06	21.67
YOLO 608		26.01	44.54	4.55	61	31.76	13.02	22.67	12.35	9.93	39.99	26.58
SSD300 [16]	ResNet-101	5.5	9.1	0	25.5	6.1	4.5	0	4.5	9.1	18.2	8.25
SSD300 [16]	VGG16	9.1	17.1	0	26.1	9.1	9.1	0	4.5	0	16.7	9.16
SSD512 [16]	VGG16	9.1	17.1	0	43	9.1	9.1	9.1	9.1	0	7.6	11.32
RetinaNet	ResNet-50-FPN	30.7	49.3	2	65.5	21.3	16.1	8.5	12.9	1	25.7	23.3
RetinaNet	ResNet-101-FPN	30.6	48.7	7.1	64.7	20	15.9	11.8	10.7	2.9	38.7	25.1
RetinaNet	ResNeXT-101-32  8d-FPN	35.5	55	12.1	66.5	23.9	18.4	9.8	16.2	9.4	53.7	30
RetinaNet	ResNeXT-101-64  4d-FPN	31.4	50.2	8.9	66.3	20.8	15.3	9.4	14	2.2	32.4	25.1
R-CNN [13]	RPN prop. + VGG16	31.9	31.3	4.2	56.8	31.1	9.3	14.2	16.4	23.4	29.4	24.8
R-CNN [13]	Alexnet, 7, 300 pro	32.4	27.2	5.1	56.9	28	9.8	13.6	12.4	17.9	35.6	23.9
R-CNN [13]	VGG16, 7, 300 pro	37.3	30.3	7.2	60.6	41.5	15.8	21.5	13.7	22	33.3	28.4
R-CNN [13]	ContextNet (Alexnet, 7)	32.7	26.8	4.6	56.4	26.3	9.9	12.9	12.2	18.7	34	23.5
Fast RCNN	ResNet-50-C4	32.4	46.3	6.5	65.8	38.3	20.1	25.3	16.6	14.1	52	31.7
Fast RCNN	ResNet-50-FPN	37.4	47.3	7.3	68.9	46.7	21	32.1	17.1	9.3	45.9	33.3
Fast RCNN	ResNet-101-FPN	39.3	50.3	10.6	68.3	47.1	20.4	33.3	18.6	15.4	51.4	35.5
Fast RCNN	ResNeXT-101-32  8d-FPN	47.5	54.8	10.3	71.8	54	21.4	34.4	21.7	17.7	53.5	38.7
Fast RCNN	ResNeXT-101-64  4d-FPN	45.4	55.7	10.9	72.5	53.3	24	36.9	22.9	16	58.1	39.6
Faster R-CNN [16]	VGG16	23.76	37.65	8.03	54	16.16	11.88	15.12	9.1	6.25	37.29	21.92
Faster RCNN	ResNet-50-C4	32.2	44.6	6.6	65.9	35.2	17.5	25.7	19.6	13.7	40	30.1
Faster RCNN	ResNet-50-FPN	35.7	49.9	7.3	68.4	48.9	18.8	29.6	14.7	11.4	53.3	33.8
Faster RCNN	ResNet-101-FPN	39.8	49.2	4.9	68.2	47	18.5	29.7	14	12.9	52.2	33.7
Faster RCNN	ResNeXT-101-32  8d-FPN	49.8	56.6	11.4	72.1	56.3	23.2	37	20.8	18.8	58.7	40.5
Faster RCNN	ResNeXT-101-64  4d-FPN	49.6	58.6	12.2	72.5	54.5	23.2	36.9	20.8	20.1	63.1	41.2

The values in bold represent the best in one-stage methods, and the ones in italics represent the highest in two-stage methods.