Updated paper

Update Notes - Main Changes:

• Adding new experiments on MPII dataset. Our model achieves superior results on MPII Val and test sets!

In the previous version, I mistook the Ubody metric for the Total metric. So the results on MPII test set are:
&Head & Shoulder & Elbow & Wrist & Hip & Knee & Ankle & UBody & Total \
& 98.6 & 97.4 & 93.9 & 90.2 & 92.9 & 91.3 & 88.0 & 93.9 & 93.5

• Improving visualization results in Figure 6, Figure 7, and figures in Appendix:
• • Comparisons between TransPose-R-A4 and TransPose-H-A4 (original: TransPose-R-A4 vs TransPose-H-S)
• • Visualizing the dependency area with a threshold value to filter lower attention scores

This paper is accepted by ICCV 2021. We will update a new arxiv preprint soon.

Yaml

Initial commit

[ImageNet]: pretrained weights on ImageNet: ResNet and HRNet

We release the ResNet-50 (resnet50-19c8e357.pth), HRNet-W32 (hrnet_w32-36af842e.pth), and HRNet-W48 (hrnet_w48-8ef0771d.pth) networks pretrained on ImageNet. Note that only the initial several layers/blocks of them are taken as the convolutional blocks of TransPose models.

[COCO]: pretrained weights on COCO dataset

We release the weights of TransPose-R-A3, TransPose-R-A4, TransPose-H-S, TransPose-H-A4, and TransPose-H-A6 models that have been trained on COCO train2017 dataset. They can either be stored locally in the github repo, or loadable by the function torch.hub.load_state_dict_from_url().