Reproduce “SparseSurf: Sparse-View 3D Gaussian Splatting for Surface Reconstruction” 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

Meiying Gu, Jiawei Zhang, Jiahe Li, Xiaohan Yu, Haonan Luo, Jin Zheng, Xiao Bai

SparseSurf introduces a 3D Gaussian Splatting framework that reconstructs detailed surfaces and generates high-fidelity novel views from sparse input images. This is achieved by integrating iteratively updated stereo-based metric geometric priors and multi-view feature consistency using both training and pseudo-unseen views, leading to state-of-the-art performance on standard benchmarks.

Key insights

  • Iteratively updating stereo-derived metric depth and normal priors provides more reliable geometric supervision than static, noisy monocular priors, crucial for sparse-view accuracy.
  • Enforcing multi-view feature consistency using a pre-trained feature extractor, applied to both real training views and synthesized pseudo-views, effectively regularizes flattened Gaussians and mitigates overfitting.
  • A robust balance between high-quality novel view synthesis and accurate surface reconstruction can be achieved in sparse-view 3DGS by integrating reliable geometric constraints with comprehensive feature-level consistency.

Summary by alphaXiv.

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