Reproduce “3D Neural Edge 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
Lei Li, Songyou Peng, Zehao Yu, Shaohui Liu, Rémi Pautrat, Xiaochuan Yin, Marc Pollefeys
The authors introduce EMAP, which reconstructs 3D edge maps containing straight segments and curves from multi-view photographs without matching points across images. It learns a 3D UDF, the distance to the nearest edge, then uses edge locations and directions to extract geometry. On ABC-NEF with DexiNed, completeness error is 8.9 mm and F5 is 59.1%; LIMAP scores higher on F5.
Key insights
- Unsigned Distance Functions (UDFs) are inherently more suitable than Signed Distance Functions (SDFs) for implicitly representing open, non-watertight geometric structures like 3D edges.
- An unbiased neural rendering formulation is crucial for accurately learning and projecting 3D edge geometry from 2D observations, resolving inaccuracies observed in prior density-based neural edge fields.
- Robust parametric extraction of 3D lines and curves can be achieved by leveraging the learned UDF field's distance and gradient information, enabling precise and scalable primitive fitting without tedious manual tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.