Research map / 3D vision / NeRF and radiance fields
Global and Hierarchical Geometry Consistency Priors for Few-shot NeRFs in Indoor Scenes
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 P2NeRF, which warms up few-shot NeRF geometry with matching-based triangulated keypoint depths and adds monocular-depth group ranking and ray-weight mask losses for 360° indoor scenes.
Problem
Few-shot NeRFs degrade on 360° outward-facing indoor scenes with large viewpoint variations, producing poor novel views and depth; existing monocular-depth priors fail to guarantee geometry consistency.
Approach
Combines a global geometry prior from LoFTR-matched keypoints triangulated into coarse depths used for early warm-up, with hierarchical priors from DPT monocular depth: a group-wise depth ranking loss across percentile groups and a ray weight mask regularization that pushes weights to correct depth intervals.
Key results
- On ScanNet, P2NeRF reaches 21.03 PSNR, 0.719 SSIM, 0.209 LPIPS and 0.213 depth RMSE, the best among compared methods.
- On Replica, P2NeRF achieves 31.15 PSNR and 0.061 LPIPS, the best novel view synthesis results, while SparseNeRF has the best depth RMSE (0.151 vs 0.185).
- Ablation on ScanNet shows geometric warm-up substantially lowers depth RMSE (0.722 to 0.376) and group depth ranking outperforms local depth ranking (0.230 to 0.213 RMSE).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add P2NeRF global and hierarchical geometry priors | Mip-NeRF · ScanNet | PSNR: 20.85 -> 21.03 | Helped |
| add P2NeRF global and hierarchical geometry priors | Mip-NeRF · Replica | PSNR: 29.86 -> 31.15 | Helped |
| add matching-based geometry warm-up (ablation) | Mip-NeRF few-shot training · ScanNet | RMSE: 0.722 -> 0.376 | Helped |
| remove geometry warm-up (ablation) | Mip-NeRF with group ranking and ray weight mask · ScanNet | RMSE: 0.376 -> 0.976 | Hurt |
| add group depth ranking (ablation) | P2NeRF with warm-up and ray weight mask · ScanNet | RMSE: 0.238 -> 0.213 | Helped |
| add ray weight mask regularization (ablation) | P2NeRF with warm-up and group depth ranking · ScanNet | PSNR: 20.64 -> 21.03 | Helped |
| add ray weight mask regularization (ablation) | P2NeRF with warm-up and group depth ranking · ScanNet | RMSE: 0.204 -> 0.213 | Hurt |
| add geometric warm-up (ablation) | SparseNeRF · ScanNet | RMSE: 0.991 -> 0.234 | Helped |
| use matching-based point cloud instead of SfM point cloud for warm-up (ablation) | P2NeRF · ScanNet | RMSE: 0.409 -> 0.376 | Helped |
| replace local depth ranking with group depth ranking (ablation) | P2NeRF · ScanNet | RMSE: 0.230 -> 0.213 | Helped |
| add group depth ranking without a good initial geometry (ablation) | Mip-NeRF sparse-view training · ScanNet | PSNR: 18.71 -> 15.99 | Hurt |
Surprising findings
- Applying the hierarchical ranking and mask losses without warm-up causes heavy degradation (PSNR 15.99), so the ranking loss needs a good initial geometry.
- Matching-based triangulated warm-up gives lower depth RMSE than an SfM point cloud warm-up (0.376 vs 0.409).
Limitations
- Triangulated point cloud is very coarse, so large depth errors can distort the geometry during warm-up.
- Monocular depth estimates are affected by texture and illumination and can produce erroneous surfaces that propagate into rendering depth.
- The ray weight mask loss has an optimization blind spot in extreme ordering cases where the rendering depth is far from ground truth.
Experimental setup
Datasets: ScanNet Replica
Benchmarks: ScanNet Replica
Metrics: PSNR SSIM LPIPS depth RMSE
Baselines: DS-NeRF InfoNeRF RegNeRF FreeNeRF SCADE SparseNeRF
Builds on: Mip-NeRF DPT LoFTR SparseNeRF DS-NeRF
Closest papers by meaning
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.91)
- NeRFVS: Neural Radiance Fields for Free View Synthesis via Geometry Scaffolds (CVPR 2023 · similarity 0.90)
- FrugalNeRF: Fast Convergence for Extreme Few-shot Novel View Synthesis without Learned Priors (CVPR 2025 · similarity 0.88)
- Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video (CVPR 2025 · similarity 0.88)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.87)
- Mask-Based Modeling for Neural Radiance Fields (ICLR 2024 · similarity 0.87)
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.