Reproduce “Cross-View Completion Models are Zero-shot Correspondence Estimators” 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
Honggyu An, Jinhyeon Kim, Seonghoon Park, Jaewoo Jung, Jisang Han, Sunghwan Hong, Seungryong Kim
Research from KAIST, Korea University, and Samsung Electronics identified that the cross-attention maps within cross-view completion (CVC) models inherently capture precise dense geometric correspondences. Directly leveraging these maps enabled state-of-the-art zero-shot performance in dense matching and improved robustness in multi-frame depth estimation, outperforming prior methods.
Key insights
- The cross-attention map within cross-view completion (CVC) models inherently functions as a learned 'full cost volume' that captures precise dense geometric correspondences.
- Directly leveraging this cross-attention map, rather than relying on encoder or decoder features, enables superior zero-shot performance for geometric matching tasks.
- The learned cross-attention cost volume offers enhanced robustness to dynamic objects and image noise in multi-frame depth estimation compared to methods based on epipolar geometry.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.