search

HuangZhiChao95 / FewShotMotionTransfer

Pytorch implementation of our paper "Few-Shot Human Motion Transfer by Personalized Geometry and Texture Modeling"
Apache License 2.0
71 stars 10 forks source link

Scripts about generating the keypoint, densepose and foreground segmentation of the human #4

Closed WinstonDeng closed 2 years ago

WinstonDeng commented 2 years ago

Hi, thanks for your great job! Could you share the scripts about generating the keypoint, densepose and foreground segmentation of the human? It could be helpful if you share more guidance about details and data formats.

HuangZhiChao95 commented 2 years ago

Thanks for your attention to our work.

The generation of keypoints, densepose needs the package of Openpose and Densepose. They require a lot of dependencies and their installation may vary for different systems. You may use the docker to install them and it would avoid plenty of dependency issues.

I perform foreground segmentation by the commercial system in Huya. I am not authorized to release it.

The data format of the keypoints follows the format of Openpose. The "image_list.txt" contains only the filename without suffix, i.e. do not include "jpg" or "png" in the image_list.txt. You may find details for data processing in line 86-95 in DataSet.py.

WinstonDeng commented 2 years ago

Thanks for your guidance! Another request is that, could you share the pre-trained deeplabv3 model for segmentation? Because the download link in Human Segmentation Repo is broken and their authors don't reply. My email is 842237857@qq.com. Thanks!