Research map / 3D vision / NeRF and radiance fields
Evaluate Geometry of Radiance Fields with Low-Frequency Color Prior
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 Inverse Mean Residual Color (IMRC), a closed-form metric that scores radiance-field density geometry using only observation images by checking how well the color field is approximated by low-frequency spherical harmonics.
Problem
Evaluating the geometry (density field) of radiance fields is hard because ground-truth geometry is rarely available; Chamfer Distance requires expensive point-cloud scans and PSNR only measures rendered appearance.
Approach
Low-frequency color prior: for a good density field, the color on surfaces is low-frequency, so the transmittance-density-weighted residual after approximating observed colors with low-degree spherical harmonics is small; IMRC is the decibel transform of this mean residual color.
Key results
- On DTU, IMRC and Chamfer Distance and user ranking agree on most pairs: IMRC conflicts with UserRank on only 2 of 90 pairs, versus 11 for CD.
- Average IMRC on DTU ranks JaxNeRF 18.71, DVGO 18.19, Plenoxels 15.88, TensoRF 9.80, consistent with CD and user rank.
- Adding distortion loss to DVGO raises IMRC on all three datasets (DTU 18.19 to 18.47, NeRF Synthetic 17.25 to 17.94, LLFF 18.87 to 20.76).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add distortion loss to DVGO (ablation) | DVGO · DTU | IMRC: 18.19 -> 18.47 | Helped |
| Add distortion loss to DVGO (ablation) | DVGO · NeRF Synthetic | IMRC: 17.25 -> 17.94 | Helped |
| Add distortion loss to DVGO (ablation) | DVGO · LLFF | IMRC: 18.87 -> 20.76 | Helped |
| Increase SH degree from 0 to 3 (ablation) | IMRC metric · DTU | IMRC: 13.60 -> 16.28 | Helped |
| Increase density volume resolution from 64³ to 768³ (ablation) | IMRC metric · DTU | IMRC: 15.09 -> 15.67 | Helped |
| Increase density volume resolution from 256³ to 512³ (ablation) | IMRC metric · DTU | IMRC: 15.64 -> 15.64 | No clear effect |
| Add depth supervision (ablation) | DVGO · NeRF Synthetic | IMRC: 17.96 -> 19.49 | Helped |
| Use IMRC rather than PSNR to compare density-field geometry | radiance-field evaluation · DTU | geometry ranking agreement: null | Helped |
Surprising findings
- PSNR ranking does not match geometry quality: the paper shows TensoRF can have high PSNR but poor geometry, and DVGO has the top PSNR in some cases while its geometry is qualitatively worse.
- CD can be erroneously better due to object masks that discard floaters outside the mask.
Limitations
- Fails for purely specular or highly reflective surfaces (e.g., mirrors) where the low-frequency color prior does not hold.
- A nearly empty density field with a few well-observed points can receive a high IMRC, so it should be combined with PSNR/SSIM.
Experimental setup
Datasets: DTU NeRF Synthetic LLFF
Benchmarks: DTU NeRF Synthetic LLFF
Metrics: IMRC PSNR Chamfer Distance UserRank MSE of depth MAE of normals
Baselines: JaxNeRF Plenoxels DVGO TensoRF Chamfer Distance PSNR
Builds on: NeRF Plenoxels DVGO TensoRF distortion loss
Closest papers by meaning
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.86)
- RNb-NeuS: Reflectance and Normal-based Multi-View 3D Reconstruction (CVPR 2024 · similarity 0.86)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.85)
- DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Normalization (CVPR 2024 · similarity 0.84)
- Geometry Field Splatting with Gaussian Surfels (CVPR 2025 · similarity 0.84)
- Factored-NeuS: Reconstructing Surfaces, Illumination, and Materials of Possibly Glossy Objects (CVPR 2025 · similarity 0.84)
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.