Deep Ensemble Learning for Human Action Recognition in Still Images

Table 3

Results for NCNN in the 1.5x cropped Willow action dataset.

Algorithm	Sensitivity for each class							Overall
	Inter.W.C.	Photog.	P.Music	R.Bike	R.Horse	Running	Walking	Acc	Loss
SURF	0.03	0.01	0.01	0.01	0.02	0.01	0.01	0.1406	NA
BOF	0.46	0.32	0.44	0.44	0.46	0.40	0.44	0.4234	NA
PBOF	0.49	0.34	0.47	0.48	0.58	0.47	0.33	0.4392	NA
GIST	0.59	0.36	0.31	0.45	0.49	0.43	0.30	0.3934	NA
RF	0.59	0.32	0.27	0.45	0.39	0.35	0.30	0.3618	NA
GBM	0.46	0.36	0.30	0.32	0.39	0.37	0.25	0.3270	NA
Voting	0.69	0.38	0.32	0.42	0.46	0.31	0.30	0.3791	NA
VGG16	0.62	0.19	0.73	0.88	0.46	0.59	0.41	0.5877	1.2222
VGG16_NCNN	0.74	0.53	0.76	0.82	0.6	0.56	0.33	0.6209	1.1855
VGG19	0.72	0.36	0.7	0.81	0.67	0.7	0.38	0.6209	1.1917
VGG19_NCNN	0.69	0.31	0.79	0.82	0.75	0.64	0.36	0.6272	1.0313
ResNet50	0.87	0.22	0.52	0.81	0.95	0.48	0.5	0.5987	1.1399
ResNet50_NCNN	0.85	0.16	0.67	0.87	0.77	0.23	0.74	0.6288	0.9666