Reproduce “InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any 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
Yongzhen Hu, Hao Yu, Jiawei Wang, Xianming Zeng, Jianyun Xu, Xingliang Liu, Ruizhen Hu, Xiaowei Zhou and 2 more
InfiniSplat v2 improves feed-forward novel-view rendering by keeping geometry-derived Gaussian centers fixed while learning opacity, shape, and color. Its training aligns cameras and emphasizes larger-angle target views, allowing attributes to improve renderings without moving points. Across six datasets, pose-conditioned mean LPIPS is 0.2472 and DISTS is 0.1163; severe errors in predicted geometry remain uncorrected.
Key insights
- In this setup, preserving geometry predictions while improving Gaussian attributes performs better than fine-tuning geometry or adding learned point-position offsets.
- Training with aligned ground-truth camera poses and large-angle target views helps address coordinate mismatch and constrains Gaussian shape beyond nearby views.
- Opacity learned from image reconstruction can suppress some geometrically inconsistent points: pruning at 0.01 removed 51.8% of inconsistent points while retaining 93.3% of surface-consistent points.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.