Research map / 3D vision / NeRF and radiance fields
Pano-NeRF: Synthesizing High Dynamic Range Novel Views with Geometry from Sparse Low Dynamic Range Panoramic Images
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 irradiance fields integrated with radiance fields (Pano-NeRF) to recover geometry and reconstruct HDR novel views from sparse LDR panoramic images.
Problem
NeRF on sparse LDR panoramic images yields under-constrained geometry and cannot reconstruct HDR radiance from LDR inputs captured at fixed exposure.
Approach
Model each panoramic pixel's radiance as both a signal and a light source; add irradiance fields that compute outgoing radiance via surface rendering over incident light rays sampled from the radiance field, jointly optimized with Mip-NeRF, increasing observation counts for geometry and enabling HDR recovery via irradiance-radiance attenuation.
Key results
- Pano-NeRF reaches depth RMSE 0.72 and normal MAE 29.03 on synthetic data, versus 0.77 for RegNeRF (best baseline depth).
- On HDR reconstruction on input views, Pano-NeRF achieves PU-PSNR 43.47 vs 41.65 for Mip-NeRF+Rv and 33.06 for PanoHDR-NeRF.
- Novel view synthesis PSNR on synthetic data is 23.10 vs 22.22 for RegNeRF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add irradiance fields to Mip-NeRF (ablation) | Mip-NeRF · Evaluation scenes | D-RMSE: 1.0576 -> 0.8565 | Helped |
| Add irradiance fields to Mip-NeRF (ablation) | Mip-NeRF · Evaluation scenes | HDR-VDP3: 7.1165 -> 7.5852 | Helped |
| Add irradiance fields to Mip-NeRF (ablation) | Mip-NeRF · Evaluation scenes | N-MAE: 65.4398 -> 56.6353 | Helped |
| Add geometry prior Rv to Mip-NeRF (ablation) | Mip-NeRF · Evaluation scenes | N-MAE: 65.4398 -> 36.8086 | Helped |
| Add geometry prior Rv to Mip-NeRF (ablation) | Mip-NeRF · Evaluation scenes | D-RMSE: 1.0576 -> 1.0524 | No clear effect |
| Add geometry prior Rv to Mip-NeRF (ablation) | Mip-NeRF · Evaluation scenes | HDR-VDP3: 7.1165 -> 7.2328 | Helped |
| Add geometry prior Rv to Mip-NeRF+Irr (ablation) | Mip-NeRF+Irr · Evaluation scenes | D-RMSE: 0.8565 -> 0.8146 | Helped |
| Add geometry prior Rv to Mip-NeRF+Irr (ablation) | Mip-NeRF+Irr · Evaluation scenes | HDR-VDP3: 7.5852 -> 7.6693 | Helped |
| Add geometry prior Rv to Mip-NeRF+Irr (ablation) | Mip-NeRF+Irr · Evaluation scenes | N-MAE: 56.6353 -> 30.4295 | Helped |
| Add irradiance fields and geometry prior Rv (ablation) | Mip-NeRF · Evaluation scenes | D-RMSE: 1.0576 -> 0.8146 | Helped |
| Add irradiance fields and geometry prior Rv (ablation) | Mip-NeRF · Evaluation scenes | HDR-VDP3: 7.1165 -> 7.6693 | Helped |
Surprising findings
- Applying the Rv geometry prior alone to Mip-NeRF improves normals but does not improve depth, and can severely degrade novel views.
Limitations
- Diffuse (Lambertian) BRDF assumption limits non-Lambertian surfaces.
- Geometry prior and joint optimization depend on radiance field quality; irradiance estimated from radiance fields.
Experimental setup
Datasets: Replica synthetic scenes rendered from 5 3D models real captured HDR panoramas
Benchmarks: Replica synthetic panoramic scenes real captured HDR panoramas
Metrics: RMSE MAE PSNR SSIM LPIPS PU-PSNR PU-SSIM HDR-VDP3
Baselines: Omnifusion IndoorPanoDepth Mip-NeRF RegNeRF FreeNeRF LANet PanoHDR-NeRF InvIndoor Lighthouse LRG360
Builds on: Mip-NeRF NeRF Ref-NeRF
Closest papers by meaning
- ESR-NeRF: Emissive Source Reconstruction Using LDR Multi-view Images (CVPR 2024 · similarity 0.89)
- EvHDR-NeRF: Building High Dynamic Range Radiance Fields with Single Exposure Images and Events (AAAI 2025 · similarity 0.89)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.88)
- Optimize the Unseen - Fast NeRF Cleanup with Free Space Prior (NeurIPS 2025 · similarity 0.88)
- BLiRF: Bandlimited Radiance Fields for Dynamic Scene Modeling (AAAI 2024 · similarity 0.88)
- LuSh-NeRF: Lighting up and Sharpening NeRFs for Low-light Scenes (NeurIPS 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.