An Evaluation of Deep Learning Methods for Small Object Detection

Table 2

The parameters of models.

Method	Momentum	Decay	Gamma	Learning_rate	Batch_size	Training_days	Stepsize
YOLOv2 [16]	0.9	0.0005		0.001	8	5	25000
YOLOv3	0.9	0.0005		0.001	32	3–4	25000
SSD300 [16]	0.9	0.0005	0.1	0.000004	12	9	40000, 80000
SSD512 [16]	0.9	0.0005	0.1	0.000004	12	12	100000, 120000
RetinaNet	0.9	0.0005	0.1	0.001	64	4->12 h	25000, 35000
Fast RCNN	0.9	0.0005	0.1	0.001	64	4->12 h	25000, 35000
Faster RCNN	0.9	0.0005	0.1	0.001	64	4->12 h	25000, 35000