Reproduce “DISN: Deep Implicit Surface Network for High-quality Single-view 3D Reconstruction” 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
Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, Ulrich Neumann
DISN, from USC and Adobe Research, reconstructs high-quality, detail-rich 3D meshes from single 2D images by predicting Signed Distance Fields. The method excels at capturing intricate details like holes and thin structures, outperforming prior approaches in quantitative metrics such as Earth Mover's Distance and Intersection over Union.
Key insights
- Local image features, extracted by projecting 3D query points onto the 2D image plane, are critical for reconstructing fine-grained details in 3D shapes from single views.
- Regressing continuous Signed Distance Function (SDF) values with a weighted L1 loss provides a more accurate and robust training signal than binary occupancy classification.
- Accurate camera pose estimation is essential for effectively linking 3D space with 2D image evidence, enabling the extraction of relevant local features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.