Research map / 3D vision / Gaussian splatting
FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FewViewGS, a 3D Gaussian Splatting method for few-shot novel view synthesis that enforces matching-based geometry, color, and semantic consistency on novel views sampled between training views, plus a locality-preserving regularizer, using a multi-stage training schedule and no pretrained depth or diffusion priors.
Problem
Gaussian Splatting overfits and degrades when only a few (3-view) training images are available, producing poor novel view rendering; prior few-shot 3DGS methods rely on pre-trained depth or diffusion priors that are sensitive to domain shift or slow.
Approach
Sample novel poses between pairs of training views, use robust feature matches (RoMa) between the training views warped into the novel view as supervision with agreement masking, color-gradient weighting and min-over-pair losses (geometry, color, VGG semantic), staged as pre-training, intermediate and tuning, plus a KNN-based locality-preserving color regularizer.
Key results
- On DTU with 3 views, FewViewGS reaches 19.74 PSNR / 0.861 SSIM / 0.127 LPIPS, versus 16.94 PSNR for 3DGS, and the paper reports SoTA on LLFF at 20.54 PSNR / 0.693 SSIM / 0.214 LPIPS.
- On the Blender dataset with 8 views, FewViewGS reports 25.550 PSNR, above DNGaussian at 24.305 and FreeNeRF at 24.259.
- Ablation on DTU shows the locality regularizer raises the full model to 19.13 PSNR, and replacing RoMa matching with SIFT drops PSNR to 16.04.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FewViewGS instead of 3DGS | FewViewGS · DTU, 3 views | SSIM: 0.816 -> 0.861 | Helped |
| Use FewViewGS instead of 3DGS | FewViewGS · LLFF, 3 views | PSNR: 19.48 -> 20.54 | Helped |
| Add locality-preserving regularization to 3DGS (ablation) | 3DGS · DTU, 3 views, random initialization | PSNR: 15.04 -> 15.64 | Helped |
| Add geometry consistency loss alone (ablation) | FewViewGS · DTU, 3 views, random initialization | PSNR: 15.04 -> 18.17 | Helped |
| Add color consistency loss alone (ablation) | FewViewGS · DTU, 3 views, random initialization | PSNR: 15.04 -> 18.50 | Helped |
| Add semantic consistency loss alone (ablation) | FewViewGS · DTU, 3 views, random initialization | PSNR: 15.04 -> 15.52 | Helped |
| Use VGG semantic features instead of DINOv2 or CLIP (ablation) | FewViewGS · DTU, 3 views, random initialization | PSNR: 18.71; 17.86; 18.95 | Mixed |
| Remove the min operation from consistency losses (ablation) | FewViewGS · DTU, 3 views, random initialization | PSNR: 18.95 -> 18.28 | Hurt |
| Replace multi-stage training with single-stage training (ablation) | FewViewGS · DTU, 3 views, random initialization | PSNR: 18.95 -> 18.15 | Hurt |
| Remove pairwise feature matching and use one known view (ablation) | FewViewGS · DTU, 3 views, random initialization | PSNR: 18.95 -> 16.05 | Hurt |
| Replace RoMa matching with SIFT (ablation) | FewViewGS · DTU, 3 views, random initialization | PSNR: 18.95 -> 16.04 | Hurt |
| Set geometry consistency weight alpha to 0.2 (ablation) | FewViewGS · DTU, 3 views | PSNR: 19.13 -> 18.50 | Hurt |
Surprising findings
- DINOv2 and CLIP semantic features performed worse than the low-level VGG16 features for the semantic loss, attributed to early down-sampling losing detail.
- On DTU the method falls short of FreeNeRF in PSNR, which the authors attribute to FreeNeRF hallucinating missing regions while explicit Gaussians cannot.
Limitations
- Struggles to render texture-rich regions accurately, especially when novel views diverge significantly from input views.
- Quality depends on the feature matcher; a more precise matcher producing denser pairs would help, and SIFT matching fails due to few matches.
Experimental setup
Datasets: DTU LLFF Blender
Benchmarks: DTU LLFF Blender
Metrics: PSNR SSIM LPIPS
Baselines: NeRF SRF PixelNeRF MVSNeRF Mip-NeRF DietNeRF RegNeRF FreeNeRF SparseNeRF 3DGS DNGaussian FSGS
Builds on: 3D Gaussian Splatting RoMa FSGS DNGaussian DietNeRF
Closest papers by meaning
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.94)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.93)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.93)
- DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting (CVPR 2025 · similarity 0.93)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.92)
- HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.