hustvl / EfficientPose

EfficientPose

The inference code of the paper EfficientPose: Efficient Human Pose Estimation with Neural Architecture Search

We propose an efficient framework targeted at human pose estimation including two parts, the efficient backbone and the efficient head. We use NAS(Neural architecture search) technology to obtain lightweight backbone. For the efficient head, we slim the transposed convolutions and propose a spatial information correction module to promote the performance of the final prediction.

Requirements

• pytorch 1.0.1+
• python 3.5+

Main Results

• model & training log can be found in our model_zoo

Results on MPII val

Arch	Pretrain	Params	GFLOPs	PCKh@0.5
SimpleBaseline-R50	Y	34.0M	12.0	88.5
EfficientPose-A	N	1.3M	0.7	88.1
SimpleBaseline-R101	Y	52.0M	19.1	89.1
EfficientPose-B	N	3.3M	1.5	89.3
SimpleBaseline-R152	Y	68.6M	21.0	89.6
EfficientPose-C	N	5.0M	2.0	89.5

• Flip test is used. Input size is 256x256.
• The FNA previously proposed by our group is used.

Results on COCO val2017

Arch	Pretrain	GFLOPs	AP
EfficientPose-A	N	0.5	0.665
EfficientPose-B	N	1.1	0.711
EfficientPose-C	N	1.6	0.713

Results on COCO test2017

Acknowledgement

We thank for open-source implementation of HRNet.