Reproduce “Unsupervised Adversarial Learning of 3D Human Pose from 2D Joint Locations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yasunori Kudo, Keisuke Ogaki, Yusuke Matsui, Yuri Odagiri
Researchers developed the first unsupervised method for estimating 3D human pose from single 2D images, leveraging a GAN-based framework with a rotation-projection mechanism to learn depth without any 3D training data. The approach by Kudo et al. produces visually plausible 3D poses for diverse 'in the wild' datasets, achieving 130.9 mm error on Human3.6M when using ground truth 2D inputs.
Key insights
- Geometric consistency, where a 3D pose when rotated and re-projected yields a plausible 2D pose, can serve as a powerful self-supervisory signal for learning 3D structure.
- Depth information necessary for 3D human pose can be effectively inferred from 2D joint locations alone, eliminating the need for 3D ground truth during training.
- A simple anatomical heuristic, such as the relative orientation of the face and shoulders, can resolve inherent z-axis ambiguities in 3D reconstruction from 2D projections.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.