Reproduce “Hyperspherical Embedding for Point Cloud Completion” 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

Junming Zhang, Haomeng Zhang, Ram Vasudevan, Matthew Johnson-Roberson

Researchers from the University of Michigan and Carnegie Mellon University developed a hyperspherical embedding module that enhances point cloud completion by constraining learned features onto a unit hypersphere, leading to more compact representations and stable training. The module consistently improved Chamfer Distance by 3% to 9% across various baseline models and boosted performance in multi-task settings on real-world datasets like GraspNet, while also mitigating optimization conflicts.

Key insights

  • Optimizing L2 normalized embeddings results in gradient updates that primarily alter the direction, not the magnitude, of the pre-normalization embedding.
  • The L2 norm of the embedding before normalization monotonically increases during training, which effectively regularizes the learning rate and enhances training stability across a wider range of rates.
  • The module leads to learning poorly conditioned weights that reduce data complexity, resulting in a more compact angular distribution of embeddings in the latent space.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.