Reproduce “MV-RoMa: From Pairwise Matching into Multi-View Track Reconstruction” 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
Jongmin Lee, Seungyeop Kang, Sungjoo Yoo
MV-RoMa uses initial pairwise matches as sparse track priors, then estimates dense correspondences from a source image to several co-visible targets together, rather than relying only on chained pairwise links. On ETH3D’s Type B full-SfM comparison, it reaches 85.88% accuracy at 1 cm versus 84.79% for RoMa plus Dense-SfM, but completeness is lower: 3.95% versus 7.38%.
Key insights
- Directly learning geometrically consistent dense correspondences across multiple views significantly outperforms chaining and refining individual pairwise matches for 3D reconstruction.
- Sparse geometric priors, represented as 'track tokens' obtained through clustering-based sampling, effectively guide a dense feature encoder to learn multi-view-aware representations.
- An efficient multi-view attention mechanism, specifically pixel-aligned attention during refinement, enables coherent information exchange across views without incurring prohibitive computational overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.