Reproduce “StreetSurf: Extending Multi-view Implicit Surface Reconstruction to Street Views” 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
Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang and 1 more
StreetSurf presents a multi-view implicit surface reconstruction technique tailored for unbounded street scenes in autonomous driving datasets, achieving high-fidelity geometry and appearance efficiently. It effectively reconstructs detailed urban environments without requiring dense LiDAR data by leveraging monocular priors.
Key insights
- Decomposing unbounded street scenes into specialized spatial regions (close-range, distant-view, sky) and modeling them with distinct networks significantly improves reconstruction quality and disentanglement.
- Tailoring scene representations with cuboid-aligned spaces and hash-grids, proportional to the long, narrow street geometry, enhances efficiency and model capacity utilization.
- Incorporating geometric priors, derived from either sparse LiDAR or general-purpose monocular depth/normal estimators, guides accurate implicit surface reconstruction, especially in data-scarce regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.