Reproduce “GAPS: Generative Active Pseudo-view Selection for Sparse-View 3D 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
Hongfei Zhu, Haochen Deng, Sitao Zhang, Ling Zhou
GAPS generates selected synthetic views with diffusion to improve sparse-view 3D Gaussian Splatting reconstructions. It scores candidate viewpoints for reconstruction benefit and generation reliability, then filters them using coverage and image-residual checks. On seven Mip-NeRF 360 scenes with 24 training views, it raises average PSNR from 22.23 to 23.03 dB versus vanilla 3DGS, though some scenes regress.
Key insights
- Generated supervision is most effective when the system chooses where it is needed and filters views that are likely to be unreliable, rather than treating all diffusion outputs equally.
- Confidence annealing enables conservative pseudo-view selection early in training and broader exploration as the reconstruction becomes more stable.
- Pseudo-view selection and Gaussian dropout address complementary issues: GAPS primarily improves perceptual consistency and coverage, while DropGaussian provides stronger protection against representation overfitting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.