Reproduce “Single-view 3D Scene Reconstruction with High-fidelity Shape and Texture” 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
Yixin Chen, Junfeng Ni, Nan Jiang, Yaowei Zhang, Yixin Zhu, Siyuan Huang
Single-view neural implicit Shape and Radiance field (SSR) reconstructs detailed 3D scenes from single images, achieving high-fidelity shape and texture simultaneously. The method significantly outperforms state-of-the-art techniques on 3D-FRONT and Pix3D datasets, demonstrating substantial gains in Chamfer Distance and F-Score while enabling novel view synthesis and 3D scene editing.
Key insights
- The tight integration of implicit shape and radiance fields allows for simultaneous high-fidelity reconstruction of both geometry and texture from a single image.
- A carefully designed two-stage learning curriculum effectively resolves shape-appearance ambiguity by first establishing a strong 3D shape prior before gradually integrating 2D photometric, depth, and normal supervision.
- Leveraging 2D geometric cues (depth and normals) during training provides essential guidance to learn fine 3D details and improve reconstruction robustness from single-view inputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.