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

DeepFashion Dataset

Results on DeepFashion dataset with ResNet backbones:

Model	Input Size	PCK@0.2	AUC	EPE	Details and Download
HRNet-w48-UDP-Upper	256x192	96.1	60.9	15.1	hrnet_deepfashion.md
HRNet-w48-UDP-Lower	256x192	97.8	76.1	8.9	hrnet_deepfashion.md
HRNet-w48-UDP-Full	256x192	98.3	67.3	11.7	hrnet_deepfashion.md
ResNet-50-Upper	256x192	95.4	57.8	16.8	resnet_deepfashion.md
ResNet-50-Lower	256x192	96.5	74.4	10.5	resnet_deepfashion.md
ResNet-50-Full	256x192	97.7	66.4	12.7	resnet_deepfashion.md

DeepFashion2 Dataset

Results on DeepFashion2 dataset

Model	Input Size	PCK@0.2	AUC	EPE	Details and Download
ResNet-50-Short-Sleeved-Shirt	256x192	0.988	0.703	10.2	res50_deepfashion2.md
ResNet-50-Long-Sleeved-Shirt	256x192	0.973	0.587	16.6	res50_deepfashion2.md
ResNet-50-Short-Sleeved-Outwear	256x192	0.966	0.408	24.0	res50_deepfashion2.md
ResNet-50-Long-Sleeved-Outwear	256x192	0.987	0.517	18.1	res50_deepfashion2.md
ResNet-50-Vest	256x192	0.981	0.643	12.7	res50_deepfashion2.md
ResNet-50-Sling	256x192	0.940	0.557	21.6	res50_deepfashion2.md
ResNet-50-Shorts	256x192	0.975	0.682	12.4	res50_deepfashion2.md
ResNet-50-Trousers	256x192	0.973	0.625	14.8	res50_deepfashion2.md
ResNet-50-Skirt	256x192	0.952	0.653	16.6	res50_deepfashion2.md
ResNet-50-Short-Sleeved-Dress	256x192	0.980	0.603	15.6	res50_deepfashion2.md
ResNet-50-Long-Sleeved-Dress	256x192	0.976	0.518	20.1	res50_deepfashion2.md
ResNet-50-Vest-Dress	256x192	0.980	0.600	16.0	res50_deepfashion2.md
ResNet-50-Sling-Dress	256x192	0.967	0.544	19.5	res50_deepfashion2.md