Research map / 3D vision / Gaussian splatting
Binocular-Guided 3D Gaussian Splatting with View Consistency for Sparse View Synthesis
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 a prior-free sparse-view 3D Gaussian Splatting method that uses binocular stereo consistency between rendered views, an opacity decay strategy, and dense matcher-based initialization.
Problem
Novel view synthesis from sparse input views with 3D Gaussian Splatting, where overfitting causes geometric inaccuracies and existing methods rely on noisy pretrained depth priors.
Approach
Self-supervised binocular stereo consistency: render a slightly translated view, use rendered depth to compute disparity, warp the rendered image back to the input view and penalize photometric error; combined with an opacity decay that prunes Gaussians far from surfaces and dense keypoint-matcher initialization.
Key results
- On LLFF with 3 views, the method reaches 21.44 PSNR / 0.751 SSIM / 0.168 LPIPS, versus 19.12 PSNR for DNGaussian and 20.31 for FSGS.
- On DTU with 3 views, the method reaches 20.71 PSNR / 0.862 SSIM / 0.111 LPIPS, versus 18.91 PSNR for DNGaussian.
- On Blender with 8 views, the method reaches 24.71 PSNR, slightly above DNGaussian (24.31), with SSIM and LPIPS comparable to the best baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed method instead of ReconFusion | LLFF · LLFF, 3-view | PSNR: 21.34 -> 21.44 | Helped |
| Use the proposed method instead of the strongest baseline | DTU · DTU, 3-view | LPIPS: 0.124 -> 0.111 | Helped |
| Use the proposed method instead of DNGaussian | Blender · Blender, 8 input views | PSNR: 24.31 -> 24.71 | Helped |
| Add dense initialization, binocular consistency loss, and opacity decay (ablation) | 3DGS · LLFF, 3-view | PSNR: 15.52 -> 21.44 | Helped |
| Add dense initialization, binocular consistency loss, and opacity decay (ablation) | 3DGS · DTU, 3-view | PSNR: 10.99 -> 20.71 | Helped |
| Add opacity decay (ablation) | Ours with dense initialization and binocular consistency · LLFF, 3-view | PSNR: 20.48 -> 21.44 | Helped |
| Add dense initialization (ablation) | Ours with binocular consistency and opacity decay · DTU, 3-view | PSNR: 15.33 -> 20.71 | Helped |
| Add binocular consistency loss (ablation) | Ours with dense initialization and opacity decay · LLFF, 3-view | PSNR: 21.13 -> 21.44 | Helped |
| Add binocular consistency loss and opacity decay (ablation) | 3DGS · Blender, 3 input views | PSNR: 22.23 -> 24.71 | Helped |
| Replace opacity decay with opacity entropy regularization (ablation) | Ours · LLFF, 3-view | PSNR: 15.75 -> 21.44 | Hurt |
| Use background masks on training views (ablation) | Ours · DTU | PSNR: 20.71 -> 22.03 | Helped |
| Set opacity decay coefficient to 0.995 rather than 1.0 (ablation) | Ours · LLFF, 3-view | PSNR: 20.21 -> 21.44 | Helped |
Surprising findings
- Ablation on DTU shows the binocular consistency loss alone can hurt PSNR (13.41 vs 10.99 baseline gain is positive, but opacity decay alone lowers DTU PSNR to 14.40), and white background regions cause white Gaussians that occlude objects.
- Entropy regularization on opacity (Sugar-style) yields 15.75 PSNR on LLFF, far below opacity decay.
Limitations
- Low-texture regions such as white DTU backgrounds yield inaccurate depth, so white Gaussians can occlude objects in novel views.
- Performance on DTU drops without dense initialization (sparse SfM init gives 15.33 PSNR), and dense init depends on a pretrained matcher.
Experimental setup
Datasets: LLFF DTU NeRF Blender Synthetic
Benchmarks: LLFF DTU NeRF Blender Synthetic
Metrics: PSNR SSIM LPIPS
Baselines: DietNeRF RegNeRF FreeNeRF SparseNeRF ReconFusion MuRF 3DGS FSGS DNGaussian
Builds on: 3D Gaussian Splatting DNGaussian FSGS PDC-Net+
Closest papers by meaning
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.91)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.91)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.91)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.90)
- DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting (CVPR 2025 · similarity 0.90)
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.