Towards 3D Human Pose Estimation in the Wild: a Weakly-supervised Approach
Code: https://github.com/~xingyizhou/pose-hg-3d.
Approach
- 2D pose estimation module and a depth regression module
- Training set: images with 3D groundtruth in the lab + images with only 2D ground truth in the wild
3D depth regression module
- Integration of 2D and 3D module
- 3D geometric constraint induced loss
- How to deal with 2D weakly-labeled data?
- => a loss induced from a geometric constraint(effective regularization for depth prediction)
- and : corresponding loss weights