Reproduce “Building Rome from a Single Image” 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
Jiraphon Yenphraphai, Fang Li, Tianshuo Xu, Depu Meng, Quentin Herau, Yihan Hu, Raymond A. Yeh, Wei Zhan
The method reconstructs a scene mesh from one photograph by varying generation-region size with depth and conditioning geometry generation on estimated surface locations and visibility cues. On seven Tanks and Temples scenes, it reports normalized Chamfer distance 1.99 versus VolFill’s 2.63, and F1 0.506 versus Lyra 2.0’s 0.388; hidden geometry is plausible completion, not observed fact.
Key insights
- Adaptive chunk scale addresses the trade-off between scene coverage, geometric detail, memory use, and generation time more effectively than uniform chunking in the reported tests.
- Explicit image-to-3D correspondence and visibility information help guide where observed geometry should be retained and where the model must complete the scene.
- Synthetic outdoor data can improve not only outdoor results but also indoor validation performance when rendered inputs and target meshes are geometrically consistent.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.