An Evaluation of Deep Learning Methods for Small Object Detection

Table 6

The comparison of consumption on subsets filtered from PASCAL VOC.

Model	Backbone	Inference time (s)	Test RAM (MiB)	Train RAM (MiB)
YOLOv3	Darknet-53	0.027	1645	4079
RetinaNet	ResNet-50-FPN	0.1	1935	4133
RetinaNet	ResNet-101-FPN	0.116	2585	5435
RetinaNet	ResNeXT-101-32  8d-FPN	0.222	3641	7723
RetinaNet	ResNeXT-101-64  4d-FPN	0.284	3561	7599
Fast RCNN	ResNet-50-C4	0.495	6371	5677
Fast RCNN	ResNet-50-FPN	0.092	2131	4387
Fast RCNN	ResNet-101-FPN	0.114	2819	5463
Fast RCNN	ResNeXT-101-32  8d-FPN	0.213	3873	4637
Fast RCNN	ResNeXT-101-64  4d-FPN	0.265	3735	4575
Faster RCNN	ResNet-50-C4	0.26	6141	5991
Faster RCNN	ResNet-50-FPN	0.1	2245	5207
Faster RCNN	ResNet-101-FPN	0.13	2855	6335
Faster RCNN	ResNeXT-101-32  8d-FPN	0.225	3943	5087
Faster RCNN	ResNeXT-101-64  4d-FPN	0.276	3885	4909