Simple and Lightweight Human Pose Estimation

Introduction

This is an official pytorch implementation of Simple and Lightweight Human Pose Estimation. The codes are developed based on the repository of HRNet.

Experimental Results

Results on MPII val set

Method	#Params	FLOPs	Head	Shoulder	Elbow	Wrist	Hip	Knee	Ankle	Mean	[email protected]
pose_resnet_501	34.0M	12.0G	96.4	95.3	89.0	83.2	88.4	84.0	79.6	88.5	34.0
lpn_50	2.9M	1.3G	96.56	95.33	88.51	83.50	88.84	84.00	79.81	88.64	34.12

Note:

• Flip test is used.
• Input size is 256x256.

Results on COCO val2017 with detector having human AP of 56.4 on COCO val2017 dataset

Method	#Params	FLOPs	AP	Ap .5	AP .75	AP (M)	AP (L)	AR	AR .5	AR .75	AR (M)	AR (L)
pose_resnet_501	34.0M	8.9G	0.704	0.886	0.783	0.671	0.772	0.763	0.929	0.834	0.721	0.824
pose_resnet_1011	53.0M	12.4G	0.714	0.893	0.793	0.681	0.781	0.771	0.934	0.840	0.730	0.832
pose_resnet_1521	68.6M	15.7G	0.720	0.893	0.798	0.687	0.789	0.778	0.934	0.846	0.736	0.839
pose_hrnet_w322	28.5M	7.1G	0.744	0.905	0.819	0.708	0.810	0.798	0.942	0.865	0.757	0.858
pose_hrnet_w482	63.6M	14.6G	0.751	0.906	0.822	0.715	0.818	0.804	0.943	0.867	0.762	0.864
lpn_50	2.9M	1.0G	0.691	0.881	0.766	0.659	0.757	0.749	0.923	0.818	0.707	0.810
lpn_101	5.3M	1.4G	0.704	0.886	0.781	0.672	0.772	0.762	0.929	0.831	0.721	0.822
lpn_152	7.4M	1.8G	0.710	0.892	0.786	0.678	0.777	0.768	0.933	0.834	0.726	0.827

Note:

• Flip test is used.
• Input size is 256x192.

Inference Speed on Intel I7-8700K CPU

Note:

• Flip test is used when testing the inference speed.
• For higher FPS, you can make the FLIP_TEST false.

Installation and Preparation

Please refer to HRNet's quick start

Test

Testing on MPII dataset using model zoo's models(GoogleDrive)

python test.py \
    --cfg experiments/mpii/lpn/lpn50_256x256_gd256x2_gc.yaml

Testing on COCO val2017 dataset using model zoo's models(GoogleDrive)

python test.py \
    --cfg experiments/coco/lpn/lpn50_256x192_gd256x2_gc.yaml

References

[1] Simple Baselines for Human Pose Estimation and Tracking

[2] Deep High-Resolution Representation Learning for Human Pose Estimation