Reproduce “RIGI: Rectifying Image-to-3D Generation Inconsistency via Uncertainty-aware Learning” 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
Jiacheng Wang, Zhedong Zheng, Wei Xu, Ping Liu
RIGI improves single-image 3D reconstruction by reducing the influence of pixels where two optimized 3D Gaussian Splatting models disagree. The models render the same sampled views, and their pixel-wise differences form an uncertainty proxy, not a judgment of which view is wrong. On 25 Google Scanned Objects, RIGI reports SSIM 0.8346 versus Hi3D’s 0.8217, but not the best PSNR.
Key insights
- Aleatoric uncertainty in generated multi-view pseudo-labels can be effectively quantified by observing discrepancies between two independently optimized 3D Gaussian models.
- Dynamically down-weighting the loss contribution from uncertain regions prevents conflicting signals from causing excessive densification or artifacts during 3D reconstruction.
- Combining diverse camera pose sampling (including top/bottom views) with progressive multi-scale rendering and uncertainty awareness is crucial for robust, high-fidelity 3D asset generation across all perspectives.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.