Reproduce “MVSAnywhere: Zero-Shot Multi-View Stereo” 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
Sergio Izquierdo, Mohamed Sayed, Michael Firman, Guillermo Garcia-Hernando, Daniyar Turmukhambetov, Javier Civera, Oisin Mac Aodha, Gabriel Brostow and 1 more
MVSAnywhere estimates depth from calibrated images without a supplied range. Geometry sets an initial search; cross-view matches and reference-image cues inform a prediction, then the system refines the search. Across five held-out datasets, average rel is 2.7 versus 3.4 for triangulated MAST3R, and τ is 77.0 versus 70.1; ETH3D favors MAST3R. Known poses and intrinsics remain required.
Key insights
- Combining monocular and multi-view features through transformers improves generalization
- Adaptive cost volumes enable handling arbitrary depth ranges without prior knowledge
- Pre-trained monocular depth estimation provides valuable complementary information
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.