Detecting Small Chinese Traffic Signs via Improved YOLOv3 Method

Table 2

Comparison table.

Names	TT 100 k		Our dataset
	mAP	FPS	mAP (%)	FPS
HOG + SVM [10]	0.1	2.2	4.9	12.4
Faster R-CNN [11]	50.3	3.6	74.9	6.20
YOLOv1 [13]	16.9	41.0	57.8	72.0
YOLOv1-tiny	11.5	54.9	38.6	168.0
YOLOv2 [14]	19.8	31.2	65.1	93.7
YOLOv3 [15]	52.9	25.6	93.1	49.2
YOLOv3 + fourth detection layer	54.2	13.3	98.6	43.9
Improved YOLOv3 (YOLOv3 + fourth detection layer + soft-NMS)	54.3	13.3	98.8	43.8