Reproduce “GeoRefine: Self-Supervised Online Depth Refinement for Accurate Dense Mapping” 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
Pan Ji, Qingan Yan, Yuxin Ma, Yi Xu
GeoRefine introduces a self-supervised online depth refinement system that continuously improves monocular depth predictions to achieve accurate, geometrically consistent dense mapping. This system integrates a learning-enhanced geometric SLAM with a deep learning depth model, yielding reductions of 2-4 times in absolute relative depth errors on datasets like EuRoC and TUM-RGBD compared to pre-trained baselines.
Key insights
- An online, self-supervised integration of robust geometric SLAM with deep learning depth models enables continuous improvement of dense 3D perception with high accuracy and geometric consistency.
- Leveraging advanced optical flow (RAFT) within a SLAM framework significantly enhances tracking robustness, which in turn provides reliable geometric priors for depth refinement.
- Employing diverse self-supervised loss functions, including direct supervision from sparse SLAM map points and an occlusion-aware depth consistency term, effectively refines dense depth predictions in an online fashion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.