Research map / 3D vision / NeRF and radiance fields
Learning with Unreliability: Fast Few-shot Voxel Radiance Fields with Relative Geometric Consistency
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 ReVoRF, a voxel-based few-shot radiance field method that exploits unreliable warped pseudo-views via a bilateral geometric consistency loss (color/density for reliable regions, relative depth for unreliable ones) plus reliability-aware voxel smoothing and learning.
Problem
Few-shot novel view synthesis with radiance fields suffers from unreliable warped pseudo novel views (disoccluded holes and artifacts); prior methods mask these regions, losing supervision and causing ambiguous geometry.
Approach
Treat unreliable disoccluded regions as useful via relative depth ordering consistency rather than absolute color supervision, combined with a bilateral loss and reliability-weighted voxel smoothing and gradient weighting for voxel grids.
Key results
- On Realistic Synthetic 360 (4 views), ReVoRF reaches 20.72 PSNR, 0.848 SSIM, 0.179 LPIPS in 7 minutes, outperforming GeCoNeRF (19.78 PSNR).
- On LLFF (3 views), ReVoRF achieves 19.26 PSNR, 0.644 SSIM, 0.316 LPIPS in 11 minutes; best SSIM and LPIPS but PSNR below SparseNeRF (19.86).
- Ablation on Realistic Synthetic 360 shows PSNR rising from 17.19 baseline to 20.72 with all losses.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ReVoRF instead of GeCoNeRF | Few-shot radiance field reconstruction · Realistic Synthetic 360°, 4-view | PSNR: 19.78 -> 20.72 | Helped |
| Use ReVoRF instead of GeCoNeRF | Few-shot radiance field reconstruction · Realistic Synthetic 360°, 4-view | SSIM: 0.880 -> 0.848 | Hurt |
| Use ReVoRF instead of GeCoNeRF | Few-shot radiance field reconstruction · Realistic Synthetic 360°, 4-view | LPIPS: 0.185 -> 0.179 | Helped |
| Use ReVoRF instead of SparseNeRF | Few-shot radiance field reconstruction · LLFF, 3-view | PSNR: 19.86 -> 19.26 | Hurt |
| Use ReVoRF instead of SparseNeRF | Few-shot radiance field reconstruction · LLFF, 3-view | SSIM: 0.624 -> 0.644 | Helped |
| Add reliable-region reconstruction loss Lrel (ablation) | DVGO baseline · Realistic Synthetic 360°, 4-view | PSNR: 17.19 -> 17.79 | Helped |
| Add unreliable-region relative depth loss Lunr (ablation) | Baseline + Lrel · Realistic Synthetic 360°, 4-view | PSNR: 17.79 -> 19.01 | Helped |
| Add depth smoothness loss Lds (ablation) | Baseline + Lrel + Lunr · Realistic Synthetic 360°, 4-view | PSNR: 19.01 -> 19.23 | Helped |
| Add reliability-aware voxel smoothing Lrs (ablation) | Baseline + Lrel + Lunr + Lds · Realistic Synthetic 360°, 4-view | PSNR: 19.23 -> 20.72 | Helped |
Surprising findings
- On LLFF, ReVoRF's PSNR (19.26) is lower than SparseNeRF (19.86) and VGOS-reported numbers differ, so it does not lead on PSNR across all settings.
Limitations
- Voxel-based methods tend to produce smoothed results, losing fine details.
- Limited applicability to complex or large-scale scenes under sparse inputs.
Experimental setup
Datasets: Realistic Synthetic 360 LLFF
Benchmarks: Realistic Synthetic 360 LLFF
Metrics: PSNR SSIM LPIPS training time
Baselines: NeRF PixelNeRF DietNeRF 3DGS InfoNeRF VGOS GeCoNeRF RegNeRF SparseNeRF SRF MVSNeRF Mip-NeRF
Builds on: DVGO VGOS GeCoNeRF InfoNeRF DPT
Closest papers by meaning
- Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering (CVPR 2025 · similarity 0.89)
- CVT-xRF: Contrastive In-Voxel Transformer for 3D Consistent Radiance Fields from Sparse Inputs (CVPR 2024 · similarity 0.89)
- FrugalNeRF: Fast Convergence for Extreme Few-shot Novel View Synthesis without Learned Priors (CVPR 2025 · similarity 0.89)
- GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction (NeurIPS 2025 · similarity 0.89)
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.87)
- DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Normalization (CVPR 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.