Designing Compact Convolutional Filters for Lightweight Human Pose Estimation

Table 3

Comparisons of results on MSCOCO test-dev2017 set. #Params and flops are calculated for the pose estimation network, and those for human detection are not included.

Method	Backbone	Input	#Params	GFLOPs
Bottom-up: key point detection and grouping
OpenPose [34]	—	—	—	—	61.8	84.9	67.5	57.1	68.2	66.5
Associative embedding [35]	—	—	—	—	65.5	86.6	72.3	60.6	72.6	70.2
PersonLab [4]	—	—	—	—	68.7	89	75.4	64.1	75.5	75.4
MultiPoseNet [36]	—	—	—	—	69.6	86.3	76.6	65.0	76.3	73.5
HigherHRNet [37]	HRNet-w32		28.6M	47.9	66.4	87.5	72.8	61.2	74.2	—
Top-down: human detection and single-person key point detection
Large network
Mask-RCNN [22]	ResNet-50-FPN	—	—	—	63.1	87.3	68.7	57.8	71.4	—
G-RMI [15]	ResNet-101		42.6M	57	64.9	85.5	71.3	62.3	70.0	69.7
IPR [27]	ResNet-101		45.0M	11	67.8	88.2	74.8	63.9	74.0	—
RMPE [38]	PyraNet [39]		28.1M	26.7	72.3	89.2	79.1	68.0	78.6	—
CPN [28]	—		—	—	72.1	91.4	80.0	68.7	77.2	78.5
SimpleBaseline [25]	ResNet-152		68.6M	35.6	73.7	91.9	81.1	70.3	80.0	79.0
Small network
MobileNetV2 [19]	MobileNetV2		9.8M	3.33	66.8	90.0	74.0	62.6	73.3	72.3
ShuffleNetV2 [33]	ShuffleNetV2		7.6M	2.87	62.9	88.5	69.4	58.9	69.3	68.9
Small HRNet [17]	HRNet-W16		1.3M	1.21	55.2	85.8	61.4	51.7	61.2	61.5
Lite-HRNet [17]	Lite-HRNet-18		1.1M	0.45	66.9	89.4	74.4	64.0	72.2	72.6
MobilePoseNet	MobileNetv3 [13]		1.5M	0.55	64.8	88.8	72.4	61.9	70.2	70.7
MobilePoseNet	MobileNetv3		1.5M	1.23	67.4	89.4	74.2	64.1	73.3	73.3