Benchmarking the Robustness of Object Detection Based on Near-Real Military Scenes

Table 3

mAP value of different algorithms on various scenes.

mAP	Run close	Slightly low light	Low light	Mist	Dense fog	Rain	Snow	Middle distance	Long distance	Oblique shot	Angle 1	Angle 2	Angle 3
Faster RCNN	0.555	0.553	0.597	0.677	0.470	0.568	0.602	0.082	0.000	0.369	0.784	0.485	0.917
RetinaNet	0.671	0.670	0.675	0.622	0.441	0.695	0.659	0.080	0.005	0.284	0.763	0.482	0.924
ATSS	0.499	0.447	0.379	0.460	0.275	0.429	0.424	0.040	0.000	0.299	0.755	0.558	0.817
FoveaBox	0.550	0.536	0.548	0.575	0.555	0.506	0.567	0.049	0.030	0.378	0.695	0.490	0.896
GFocal Loss	0.508	0.524	0.593	0.652	0.398	0.464	0.588	0.040	0.024	0.352	0.675	0.475	0.920
PAFPN	0.598	0.627	0.677	0.644	0.501	0.596	0.648	0.088	0.030	0.334	0.775	0.484	0.910
RepPoints	0.563	0.593	0.629	0.640	0.582	0.608	0.625	0.243	0.030	0.428	0.725	0.584	0.898