Reproduce “DeepGMR: Learning Latent Gaussian Mixture Models for Registration” 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

Wentao Yuan, Ben Eckart, Kihwan Kim, Varun Jampani, Dieter Fox, Jan Kautz

DeepGMR introduces a learning-based point cloud registration method that integrates deep learning for robust feature extraction with a Gaussian Mixture Model (GMM) framework. It achieves global, real-time, and highly accurate registration by learning pose-invariant point-to-GMM correspondences, demonstrating superior performance on various datasets compared to existing methods.

Key insights

  • Learning point-to-GMM component correspondences directly through a neural network enables global registration without the need for initial poses or iterative refinement, making the process non-local and more robust.
  • By formulating registration as minimizing KL-divergence between two GMMs with isotropic covariances, the complex problem is reduced to a weighted Euclidean distance minimization solvable in closed-form via Umeyama's SVD.
  • Integrating deep learning's ability to learn robust, pose-invariant features with the principled noise handling of probabilistic GMMs offers a powerful, efficient, and generalizable approach to point cloud registration.

Summary by alphaXiv.

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