Reproduce “SuperGS: Consistent and Detailed 3D Super-Resolution Scene Reconstruction via Gaussian Splatting” 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
Shiyun Xie, Zhiru Wang, Yinghao Zhu, Xu Wang, Chengwei Pan, Xiwang Dong
Researchers from Beihang University developed SuperGS, a method for 3D super-resolution scene reconstruction from low-resolution inputs using Gaussian Splatting, achieving consistent and detailed novel view synthesis. The approach improves perceptual quality by reducing the FID score by 41.7% on the LLFF dataset compared to previous methods, while maintaining high PSNR and SSIM.
Key insights
- A dedicated latent feature field effectively stores and propagates foundational low-resolution scene information, serving as a consistent base for high-resolution optimization and mitigating errors from pseudo-labels.
- Leveraging high-resolution depth maps and a multi-view voting mechanism for densification significantly improves reconstruction quality while reducing Gaussian redundancy and memory requirements by preventing overfitting.
- Modeling uncertainty through variational features allows for identifying regions of low model confidence, enabling robust anchor refinement and adaptive supervision to handle inconsistent 2D pseudo-labels effectively.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.