Designing Compact Convolutional Filters for Lightweight Human Pose Estimation

Table 1

Specification for MobilePoseNet. SE denotes whether there is a squeeze-and-excite in that block. NL denotes the type of nonlinearity used. Here, HS denotes h-swish and RE denotes relu. LPB is our proposed lightweight upsampling block. bneck is the bottleneck block in MobileNetV3. is the number of key points.

Input channel	Input size	Operator	Exp size	#out	Attention	NL
3		Conv2d	—	16	—	HS	2
16		bneck,	16	16	—	RE	1
16		bneck,	64	24	—	RE	2
24		bneck,	72	24	—	RE	1
24		bneck,	72	40	SE	RE	2
40		bneck,	120	40	SE	RE	1
40		bneck,	120	40	SE	RE	1
40		bneck,	240	80	—	HS	2
80		bneck,	200	80	—	HS	1
80		bneck,	184	80	—	HS	1
80		bneck,	184	80	—	HS	1
80		bneck,	480	112	SE	HS	1
112		bneck,	672	112	SE	HS	1
112		bneck,	672	160	SE	HS	1
160		bneck,	960	160	SE	HS	1
160		bneck,	960	160	SE	HS	1
160		LPB,	320	160	SE	RE	2
160		LPB,	320	160	SE	RE	2
160		Conv2d	—		—	RE	1