Reproduce “Self Supervised Networks for Learning Latent Space Representations of Human Body Scans and Motions” 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
Emmanuel Hartman, Nicolas Charon, Martin Bauer
Researchers from the University of Houston and Florida State University developed VariShaPE and MoGeN, two self-supervised networks that robustly encode noisy 3D human body scans into latent spaces and learn the inherent geometry of human motion. VariShaPE retrieves 1000 latent codes in about 1 second, outperforming optimization-based methods significantly, while MoGeN achieves superior motion interpolation with 0.0046 L2 vertex error, providing more natural and accurate movement representations.
Key insights
- The varifold distance gradient serves as a robust, mesh-invariant feature for self-supervised learning, allowing neural networks to process noisy and unregistered 3D scans effectively.
- Complex, non-linear human motions can be accurately modeled by learning a higher-dimensional Euclidean embedding of the latent pose space where these motions become linear trajectories.
- A unified self-supervised framework can achieve both accurate latent code retrieval from raw data and natural human motion generation/interpolation, significantly improving efficiency and robustness over traditional methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.