Reproduce “Learning Object-Compositional Neural Radiance Field for Editable Scene Rendering” 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
Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, Zhaopeng Cui
Researchers from Zhejiang University, Google, and The Chinese University of Hong Kong developed an object-compositional neural radiance field that allows for high-quality, object-level editing in cluttered real-world scenes. The system disentangles individual objects from complex environments, enabling their seamless manipulation (e.g., moving, rotating) without requiring isolated object captures, while also maintaining competitive novel view synthesis quality for static scenes.
Key insights
- Learning separate scene and object radiance fields within a two-pathway architecture enables independent manipulation while maintaining photorealism.
- A scene-guided occlusion identification strategy, particularly the 3D guard mask, effectively addresses ambiguities from 2D instance masks in cluttered scenes, preventing object fragmentation.
- The multi-view supervision and learned 3D geometric structure can refine noisy 2D instance masks into smoother and more accurate 3D segmentations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.