Table 3.

Performance results for each model

Models	Backbone	mAP1	Accuracy	Precision	Recall	F1 Score	Training duration
Faster R-CNN	ResNet-101	0.926	0.747	0.910	0.788	0.840	2 h 16m
RetinaNet	Resnet-101	0.953	0.777	0.910	0.743	0.814	1 h 42m
YOLOv3	DarkNet53	0.845	0.772	0.829	0.918	0.871	56m
SSD	VGG16	0.832	0.673	0.825	0.785	0.8	1 h 14m
Libra RCNN	Xception-101	0.920	0.740	0.890	0.784	0.833	2 h 34m
Dynamic R-CNN	Resnet-50	0.904	0.747	0.930	0.780	0.848	1 h 36m
Cascade R-CNN	Resnet-101 FPN	0.913	0.724	0.897	0.790	0.840	1 h 54m
FoveaBox	ResNet-50 FPN	0.895	0.745	0.931	0.788	0.853	1 h 5m
SABL Faster R-CNN	ResNet-101 FPN	0.894	0.732	0.921	0.780	0.844	2 h 23m
ATSS	ResNet-101 FPN	0.939	0.812	0.886	0.906	0.895	1 h 33m

ATSS, Adaptive Training Sample Selection; SABL, Side-Aware Boundary Localization.