Object Detection Based on Fast/Faster RCNN Employing Fully Convolutional Architectures

Table 3

The results of Fast RCNN using convolution-based regionwise classifier. “Inc” indicates an inception block and “Res” indicates a residual block. CRC is short for convolution-based regionwise classifier. The training time refers to the consuming time of 20 iterations.

Network	CRC	Training time	Model size	mAP (%)
GoogleNet	Inc_5a, 5b, GAP	15 s	23.2 MB	65.8
	Inc_5b, GAP	7 s		62.9
	GAP	5 s		41.2
Inception_v2	Inc_4e, 5a, 5b, GAP	17 s	39.3 MB	67.0
	Inc_5a, 5b, GAP	14 s		66.6
	Inc_5b, GAP	10 s		65.6
	GAP	7 s		42.9
Inception_v3	Inc_b, c1, c2, GAP	27 s	84.1 MB	69.8
	Inc_c1, c2, GAP	23 s		69.4
	Inc_c2, GAP	18 s		68.9
	GAP	12 s		49.7
ResNet_50	Res_5a, 5b, 5c, GAP	17 s	90.7 MB	71.5
	Res_5b, 5c, GAP	13 s		70.9
	Res_5c, GAP	10 s		69.5
	GAP	8 s		50.4