Reproduce “GSM-GS: Geometry-Constrained Single and Multi-view 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
Xiao Ren, Yu Liu, Ning An, Jian Cheng, Xin Qiao, He Kong
GSM-GS extends PGSR with region-specific single-view geometry constraints and confidence-weighted normal agreement across neighboring views. On DTU, it reports mean Chamfer distance of 0.51 mm versus PGSR’s 0.53 mm, but takes longer to train: 0.45 h versus 0.28 h. In ablations, multi-view constraints reach 0.48 mm; adding both terms does not improve further. Transparent, reflective, and thin-walled scenes remain challenging.
Key insights
- Adapting geometric constraints based on regional texture characteristics (texture-rich vs. texture-less) effectively addresses challenges in preserving high-frequency details and mitigating over-smoothing.
- Enforcing consistency of local surface normals between sampled 3D point clouds from neighboring views can reduce geometric artifacts and improve multi-view structural alignment.
- A synergistic combination of single-view adaptive and multi-view geometry-guided constraints improves both fine-grained surface reconstruction and overall scene fidelity in 3D Gaussian Splatting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.