search

Walter0807 / MotionBERT

[ICCV 2023] PyTorch Implementation of "MotionBERT: A Unified Perspective on Learning Human Motion Representations"
Apache License 2.0
1.06k stars 131 forks source link

How to plot the predict 3d keypoint to original image #53

Closed zero0kiriyu closed 1 year ago

zero0kiriyu commented 1 year ago

Seems the predict 3D keypoints are not in the same coordinate with the input 2d keypoints even with --pixel option and global version config, is it possible to project the 3D keypoints back to the image space and completely align with the original image?

Walter0807 commented 1 year ago

Using --pixel and global version config should be good, can you provide your input/output value and visualizations?

zero0kiriyu commented 1 year ago

After checking all my code, I find that I use the wrong checkpoint, which is train by the configuration without global suffix. By the way, can you provide the checkpoint of MB_ft_h36m_global.yaml?

Walter0807 commented 1 year ago

You can use the checkpoint in the inference docs which is the global version checkpoint (lite).