Reproduce “SiZeUp: Fast 3D Proxy from Aerial Images via Depth Ordinal Loss” 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
Wenjun Zhou, Yunshan Li, Qiaoyu Zhu, Weidan Xiong, Hao Zhang, Daniel Cohen-Or, Hui Huang
SiZeUp makes city-scale building proxies from calibrated aerial images by fixing extracted footprints and optimizing one height per building. Selected views and a fixed monocular depth model’s relative pixel ordering guide differentiable rendering, without dense point clouds or metric-depth regression. In four real scenes, it is 23–52× faster and ranks top two in coverage and volumetric IoU, but omits fine roof detail.
Key insights
- Relative depth ordering from monocular depth models provides a more reliable geometric cue for 3D reconstruction than ambiguous metric depth values, especially when coupled with a differentiable renderer.
- Constraining the 3D representation to simple extruded 2D footprints and focusing on optimizing only building heights enables significant computational efficiency while maintaining structural coherence for proxy models.
- Fusing multiple cues (texture, semantic, depth) from foundation models can substantially improve the accuracy and robustness of building footprint extraction from challenging oblique aerial imagery.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.