Reproduce “Dy3DGS-SLAM: Monocular 3D Gaussian Splatting SLAM for Dynamic Environments” 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

Mingrui Li, Yiming Zhou, Hongxing Zhou, Xinggang Hu, Florian Roemer, Hongyu Wang, Ahmad Osman

Dy3DGS-SLAM presents the first monocular RGB-only 3D Gaussian Splatting SLAM system capable of operating in dynamic environments. The system achieves robust camera pose estimation and high-fidelity scene reconstruction by employing a sophisticated multi-modal mask fusion strategy and dynamic-aware loss functions, resulting in superior tracking accuracy and the elimination of dynamic object artifacts compared to existing methods.

Key insights

  • A probabilistic mask fusion strategy combining optical flow and monocular depth, enhanced with K-means clustering for multi-object handling, effectively identifies dynamic regions.
  • Integrating a novel camera motion loss that accounts for scale recovery and static mask filtering in the tracking pipeline enables robust camera pose estimation in dynamic scenes.
  • Modifying 3D Gaussian Splatting rendering losses to differentiate between static and dynamic regions allows for accurate scene reconstruction while suppressing artifacts from moving objects.

Summary by alphaXiv.

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