Reproduce “FSGS: Real-Time Few-shot View Synthesis using 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

Zehao Zhu, Zhiwen Fan, Yifan Jiang, Zhangyang Wang

FSGS enables real-time, photo-realistic novel view synthesis from a minimal number of input images by enhancing 3D Gaussian Splatting. The method introduces proximity-guided densification and geometry guidance through monocular depth priors and pseudo-view synthesis, achieving state-of-the-art visual quality and significantly faster rendering speeds across various datasets compared to previous approaches.

Key insights

  • Strategic, proximity-based densification of 3D Gaussians can effectively address the challenge of sparse initialization in few-shot settings.
  • Integrating monocular depth priors and pseudo-view augmentation provides effective geometric regularization to prevent overfitting and improve scene consistency when training data is limited.
  • 3D Gaussian Splatting's explicit representation can be adapted to sparse input scenarios while retaining real-time rendering speed, bridging the gap between quality and efficiency.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.