Top-down heatmap-based pose estimation

Top-down methods divide the task into two stages: object detection, followed by single-object pose estimation given object bounding boxes. Instead of estimating keypoint coordinates directly, the pose estimator will produce heatmaps which represent the likelihood of being a keypoint, following the paradigm introduced in Simple Baselines for Human Pose Estimation and Tracking.

Results and Models

300W Dataset

Results on 300W dataset

Model	Input Size	NME_common	NME_challenge	NME_full	NME_test	Details and Download
HRNetv2-w18	256x256	2.92	5.64	3.45	4.10	hrnetv2_300w.md

AFLW Dataset

Results on AFLW dataset

Model	Input Size	NME_full	NME_frontal	Details and Download
HRNetv2-w18+Dark	256x256	1.35	1.19	hrnetv2_dark_aflw.md
HRNetv2-w18	256x256	1.41	1.27	hrnetv2_aflw.md

COCO-WholeBody-Face Dataset

Results on COCO-WholeBody-Face val set

Model	Input Size	NME	Details and Download
HRNetv2-w18+Dark	256x256	0.0513	hrnetv2_dark_coco_wholebody_face.md
SCNet-50	256x256	0.0567	scnet_coco_wholebody_face.md
HRNetv2-w18	256x256	0.0569	hrnetv2_coco_wholebody_face.md
ResNet-50	256x256	0.0582	resnet_coco_wholebody_face.md
HourglassNet	256x256	0.0587	hourglass_coco_wholebody_face.md
MobileNet-v2	256x256	0.0611	mobilenetv2_coco_wholebody_face.md

COFW Dataset

Results on COFW dataset

Model	Input Size	NME	Details and Download
HRNetv2-w18	256x256	3.48	hrnetv2_cofw.md

WFLW Dataset

Results on WFLW dataset

Model	Input Size	NME_test	NME_pose	NME_illumination	NME_occlusion	NME_blur	NME_makeup	NME_expression	Details and Download
HRNetv2-w18+Dark	256x256	3.98	6.98	3.96	4.78	4.56	3.89	4.29	hrnetv2_dark_wflw.md
HRNetv2-w18+AWing	256x256	4.02	6.94	3.97	4.78	4.59	3.87	4.28	hrnetv2_awing_wflw.md
HRNetv2-w18	256x256	4.06	6.97	3.99	4.83	4.58	3.94	4.33	hrnetv2_wflw.md