Reproduce “MVID: Feed-Forward Multi-View Intrinsic Image Decomposition” 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
Kang Du, Duotun Wang, Wanling Li, Yirui Guan, Zeyu Wang
MVID presents a feed-forward framework for multi-view intrinsic image decomposition, disentangling sparse overlapping views into consistent albedo, diffuse shading, and non-diffuse residuals. The method achieves state-of-the-art per-frame decomposition quality (e.g., 24.72 PSNR for albedo on Hypersim) and superior cross-view consistency (e.g., 26.68 A-CV PSNR on Hypersim), enabling practical applications like consistent illumination editing and specularity removal.
Key insights
- Leveraging a shared multi-view scene representation, implicitly capturing geometry and appearance, is effective for promoting cross-view consistency in intrinsic image decomposition.
- A dedicated factor query adapter mechanism is crucial for disentangling intrinsic factors by enabling factor-specific information extraction from a shared context, minimizing leakage between albedo, shading, and residual.
- Combining synthetic supervision with self-supervised RGB reconstruction on unlabeled real-world multi-view data allows feed-forward models to generalize robustly to 'in-the-wild' intrinsic decomposition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.