Reproduce “SCE-SLAM: Scale-Consistent Monocular SLAM via Scene Coordinate Embeddings” 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
Yuchen Wu, Jiahe Li, Xiaohan Yu, Lina Yu, Jin Zheng, Xiao Bai
SCE-SLAM introduces a monocular SLAM system that addresses scale drift by utilizing learned patch-level scene coordinate embeddings as a persistent geometric memory. This method achieves scale-consistent trajectory estimation with real-time performance, significantly reducing Absolute Trajectory Error on benchmarks like KITTI and Waymo.
Key insights
- Patch-level scene coordinate embeddings, recurrently updated through geometry-guided attention, can serve as an effective internal 'temporal scale memory' to mitigate scale drift in monocular SLAM.
- Combining feature similarity with 3D spatial proximity in attention mechanisms helps propagate scale consistently by enforcing the inductive bias that nearby points share related scale.
- A two-stage bootstrapping strategy and alternating optimization using both local flow and global scene coordinate constraints are effective for jointly achieving accurate local tracking and robust global scale consistency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.