Real-Time Surveillance Using Deep Learning

Table 3

Detailed results of training and testing for various object detectors.

Object detector	CNN feature extractors	No. of training images	Train-set AP	Time to train in hours	No. of testing images	Test-set AP
SqueezeNet-based FasterRCNN	Face	1805	0.22	13.86	695	0.23
	Weapon	1700	0.21	12.95	550	0.20
	Guard	1600	0.28	11.80	500	0.25
	Intruder	1550	0.20	11.56	500	0.18
GoogleNet-based FasterRCNN	Face	1805	0.96	25.08	695	0.97
	Weapon	1700	0.94	23.33	550	0.96
	Guard	1600	0.97	22.01	500	0.95
	Intruder	1550	0.80	21.64	500	0.79
ResNet-18-based FasterRCNN	Face	1805	0.97	17.01	695	0.97
	Weapon	1700	0.95	15.23	550	0.94
	Guard	1600	0.96	14.85	500	0.93
	Intruder	1550	0.79	14.66	500	0.83
ResNet-50-based FasterRCNN	Face	1805	0.97	47.01	695	0.98
	Weapon	1700	0.96	45.20	550	0.97
	Guard	1600	0.97	44.09	500	0.96
	Intruder	1550	0.81	42.77	500	0.85