search

QitaoZhao / ContextAware-PoseFormer

The project is an official implementation of our paper "A Single 2D Pose With Context is Worth Hundreds for 3D Human Pose Estimation".
65 stars 4 forks source link

can not reproduce the same setting as the paper #14

Closed ChChen666 closed 2 months ago

ChChen666 commented 2 months ago

Great Work! Following the code you provided, we can only obtain results with HRNet as the backbone of image features and CPN as the 2D detector, which is not consistent with the results in the paper. Could you please provide the pre-processed labels for the 2D detection results with HRNet?

QitaoZhao commented 2 months ago

Sorry for the confusion. Please refer to this issue https://github.com/QitaoZhao/ContextAware-PoseFormer/issues/4#issuecomment-1885569798. We used keypoint from CPN (which can be different from the feature backbone used) in our paper to keep the same setting with other papers. However, we do have pre-processed HRNet keypoint in https://drive.google.com/drive/folders/1OYKWnu_5GPLRfceD3Psf4-JZkloBodKx. You may need to adjust a few lines in https://github.com/QitaoZhao/ContextAware-PoseFormer/blob/main/ContextPose/mvn/datasets/human36m.py. Check that issue for more details!