README.md

Deep high-resolution representation learning for human pose estimation

Introduction

@inproceedings{sun2019deep,
  title={Deep high-resolution representation learning for human pose estimation},
  author={Sun, Ke and Xiao, Bin and Liu, Dong and Wang, Jingdong},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
  pages={5693--5703},
  year={2019}
}

Results and models

2d Human Pose Estimation

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

Arch	Input Size	AP	AP50	AP75	AR	AR50	ckpt	log
pose_hrnet_w32	256x192	0.746	0.904	0.819	0.799	0.942	ckpt	log
pose_hrnet_w32	384x288	0.760	0.906	0.829	0.810	0.943	ckpt	log
pose_hrnet_w48	256x192	0.756	0.907	0.825	0.806	0.942	ckpt	log
pose_hrnet_w48	384x288	0.767	0.910	0.831	0.816	0.946	ckpt	log
pose_hrnet_w32_fp161	256x192	0.746	0.905	0.88	0.800	0.943	ckpt	log

¹ Please refer to fp16/README.md for the method we use for mixed precision training.

Results on AIC val set with ground-truth bounding boxes

Arch	Input Size	AP	AP50	AP75	AR	AR50	ckpt	log
pose_hrnet_w32	256x192	0.323	0.762	0.219	0.366	0.789	ckpt	log

Results on MPII val set

Arch	Input Size	Mean	[email protected]	ckpt	log
pose_hrnet_w32	256x256	0.900	0.334	ckpt	log
pose_hrnet_w48	256x256	0.901	0.337	ckpt	log

Results on CrowdPose test with YOLOv3 human detector

Arch	Input Size	AP	AP50	AP75	AP (E)	AP (M)	AP (H)	ckpt	log
pose_hrnet_w32	256x192	0.675	0.825	0.729	0.770	0.687	0.553	ckpt	log

Results on PoseTrack2018 val with ground-truth bounding boxes

Arch	Input Size	Head	Shou	Elb	Wri	Hip	Knee	Ankl	Total	ckpt	log
pose_hrnet_w32	256x192	87.4	88.6	84.3	78.5	79.7	81.8	78.8	83.0	ckpt	log

The models are first pre-trained on COCO dataset, and then fine-tuned on PoseTrack18.

Results on PoseTrack2018 val with MMDetection pre-trained Cascade R-CNN (X-101-64x4d-FPN) human detector

Arch	Input Size	Head	Shou	Elb	Wri	Hip	Knee	Ankl	Total	ckpt	log
pose_hrnet_w32	256x192	78.0	82.9	79.5	73.8	76.9	76.6	70.2	76.9	ckpt	log

The models are first pre-trained on COCO dataset, and then fine-tuned on PoseTrack18.