An Evaluation of Deep Learning Methods for Small Object Detection

Table 5

The comparison of consumption on small object dataset.

Model	Backbone	Inference time (s)	Test RAM (MiB)	Train RAM (MiB)
YOLOv3	Darknet-53	0.0331	1825	4759
YOLOv3	ResNet-50	0.027	1285	3479
YOLOv3	ResNet-101	0.0356	1829	5383
YOLOv3	ResNet-152	0.0454	2443	7531
RetinaNet	ResNet-50-FPN	0.102	2075	4435
RetinaNet	ResNet-101-FPN	0.127	2723	5577
RetinaNet	ResNeXT-101-32  8d-FPN	0.229	3767	7863
RetinaNet	ResNeXT-101-64  4d-FPN	0.292	3719	7813
Fast RCNN	ResNet-50-C4	0.3	6449	5877
Fast RCNN	ResNet-50-FPN	0.089	2277	4455
Fast RCNN	ResNet-101-FPN	0.113	2947	5627
Fast RCNN	ResNeXT-101-32  8d-FPN	0.212	3987	4961
Fast RCNN	ResNeXT-101-64  4d-FPN	0.269	3885	4799
Faster RCNN	ResNet-50-C4	0.412	6609	6129
Faster RCNN	ResNet-50-FPN	0.101	2387	5381
Faster RCNN	ResNet-101-FPN	0.124	3001	6487
Faster RCNN	ResNeXT-101-32  8d-FPN	0.256	4027	5333
Faster RCNN	ResNeXT-101-64  4d-FPN	0.286	4003	5246