Reproduce “Scene Grounding In the Wild” 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
Tamir Cohen, Leo Segre, Shay Shomer-Chai, Shai Avidan, Hadar Averbuch-Elor
A framework for scene grounding aligns partial 3D reconstructions from "in-the-wild" imagery to a complete pseudo-synthetic reference model, addressing challenges of sparse visual overlap and domain gaps. This approach, which uses semantic features and robust optimization with 3D Gaussian Splatting, reduces average rotation error by over 50% compared to classical baselines and significantly outperforms feed-forward 3D reconstruction models in challenging scenarios.
Key insights
- Semantic features offer more robust correspondences than photometric cues for cross-domain alignment, effectively bridging the appearance gap between real-world and pseudo-synthetic imagery.
- Inverse optimization applied to 3D Gaussian Splatting (augmented with semantic features) provides an efficient and accurate method for globally aligning fragmented 3D reconstructions.
- Robust optimization techniques, particularly Least Trimmed Squares, are critical for mitigating the impact of outliers and noise inherent in real-world data and reference models, ensuring stable alignment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.