Research map / 3D vision / Gaussian splatting
NexusGS: Sparse View Synthesis with Epipolar Depth Priors in 3D Gaussian Splatting
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.
Builds a dense initial point cloud for sparse-view 3DGS by computing depth from optical flow and camera epipolar geometry, with flow-resilient depth blending and flow-filtered pruning, avoiding monocular depth priors and extra regularization.
Problem
Sparse-view novel view synthesis with 3D Gaussian Splatting struggles because sparse initial point clouds and split-based densification yield oversmoothing, overfitting, and unreliable geometry under limited supervision.
Approach
Compute per-view depth explicitly by triangulating optical-flow correspondences projected onto epipolar lines (Epipolar Depth Nexus), blend depths across views by choosing the one with minimal sensitivity to flow error (Flow-Resilient Depth Blending), and prune points with large epipolar distance (Flow-Filtered Depth Pruning), producing a dense initial point cloud so 3DGS needs no depth regulariza
Key results
- On LLFF (3 views), NexusGS reaches 21.07 PSNR, 0.738 SSIM, 0.177 LPIPS, outperforming CoR-GS (20.45 PSNR).
- On DTU (3 views), NexusGS reaches 20.21 PSNR and 0.869 SSIM, the best among compared methods.
- On Blender (8 views), NexusGS reaches 24.37 PSNR, 0.893 SSIM, slightly above DNGaussian (24.07) and CoR-GS (23.98).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NexusGS epipolar depth priors and dense point-cloud initialization instead of CoR-GS | Sparse-view novel view synthesis · LLFF (3 Views) | PSNR: 20.45 -> 21.07 | Helped |
| use NexusGS epipolar depth priors and dense point-cloud initialization instead of CoR-GS | Sparse-view novel view synthesis · DTU (3 Views) | PSNR: 19.21 -> 20.21 | Helped |
| use NexusGS epipolar depth priors and dense point-cloud initialization instead of CoR-GS | Sparse-view novel view synthesis · Mip-NeRF-360 (24 Views) | PSNR: 23.55 -> 23.86 | Helped |
| use NexusGS epipolar depth priors and dense point-cloud initialization instead of FreeNeRF | Sparse-view novel view synthesis · Blender (8 Views) | PSNR: 24.26 -> 24.37 | Helped |
| replace average depth blending with Flow-Resilient Depth Blending (ablation) | NexusGS · LLFF (3 Views) | PSNR: 20.39 -> 20.90 | Helped |
| replace nearest-view depth blending with Flow-Resilient Depth Blending (ablation) | NexusGS · LLFF (3 Views) | PSNR: 20.62 -> 20.90 | Helped |
| replace weighted-average depth blending with Flow-Resilient Depth Blending (ablation) | NexusGS · LLFF (3 Views) | PSNR: 20.59 -> 20.90 | Helped |
| add Flow-Filtered Depth Pruning to average depth blending (ablation) | NexusGS · LLFF (3 Views) | PSNR: 20.39 -> 20.51 | Helped |
| add Flow-Filtered Depth Pruning to nearest-view depth blending (ablation) | NexusGS · LLFF (3 Views) | PSNR: 20.62 -> 20.79 | Helped |
| add Flow-Filtered Depth Pruning to weighted-average depth blending (ablation) | NexusGS · LLFF (3 Views) | PSNR: 20.59 -> 20.78 | Helped |
| add Flow-Filtered Depth Pruning to Flow-Resilient Depth Blending (ablation) | NexusGS · LLFF (3 Views) | PSNR: 20.90 -> 21.07 | Helped |
Surprising findings
- Averaging depths across views performs worse than selecting depth via the flow-resilient criterion, and FSGS with only point-cloud replacement gains notably from the new point clouds.
Limitations
- Relies on optical flow estimates, which may fail due to color similarity or large pixel deviation, requiring pruning.
- Depends on known camera poses; COLMAP fails on sparse DTU inputs so random initialization is used for some baselines.
Experimental setup
Datasets: LLFF DTU Blender Mip-NeRF360
Benchmarks: LLFF DTU Mip-NeRF360 Blender
Metrics: PSNR SSIM LPIPS AVGE
Baselines: Mip-NeRF DietNeRF RegNeRF FreeNeRF SparseNeRF DS-NeRF ViP-NeRF SimpleNeRF 3DGS DNGaussian FSGS CoR-GS
Builds on: 3D Gaussian Splatting FSGS DNGaussian CoR-GS
Closest papers by meaning
- Depth-Guided Robust and Fast Point Cloud Fusion NeRF for Sparse Input Views (AAAI 2024 · similarity 0.90)
- FatesGS: Fast and Accurate Sparse-View Surface Reconstruction Using Gaussian Splatting with Depth-Feature Consistency (AAAI 2025 · similarity 0.89)
- LP-3DGS: Learning to Prune 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.89)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
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.