Reproduce “Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis” 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
Rui Peng, Wangze Xu, Luyang Tang, Liwei Liao, Jianbo Jiao, Ronggang Wang
Researchers at Peking University and the University of Birmingham developed SCGaussian, a 3D Gaussian Splatting method that ensures structural consistency for few-shot novel view synthesis by leveraging matching priors. The approach yields superior rendering quality and real-time speeds across diverse datasets, including LLFF and Tanks and Temples, often with just three input views.
Key insights
- Explicitly constraining the positions of Gaussian primitives along matched rays significantly mitigates optimization ambiguity and directly enforces 3D consistency, even with limited input data.
- Optimizing both the discrete Gaussian primitive positions and the continuous rendered geometry, guided by matching priors, is crucial for achieving high-quality, structurally consistent novel views from sparse inputs.
- A hybrid Gaussian representation (ray-based for matched regions and ordinary for others) effectively balances precise structural recovery with comprehensive scene coverage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.