Research map / 3D vision / NeRF and radiance fields
Physical-aware Neural Radiance Fields for Efficient Exposure Correction
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 PHY-NeRF, a NeRF variant with physics-inspired adaptive lighting particles in volume rendering and a feature-level global illumination consistency module, trained unsupervised to reconstruct normal-lighting novel views from low-light or overexposed multi-view images.
Problem
NeRF fails to reconstruct high-quality novel views when training images are under- or overexposed due to capture-time lighting variation and camera exposure settings.
Approach
Models illumination change inside volume rendering via per-point adaptive lighting particles (attenuating for low light, dividing for overexposure), plus a feature-level illumination embedding with consistency and mode-seeking losses to keep lighting transitions natural across views, trained without normal-light supervision.
Key results
- On overexposed scenes, PHY-NeRF reaches 21.38 PSNR / 0.764 SSIM / 0.435 LPIPS on average, versus 18.09 PSNR for Aleth-NeRF.
- On underexposed scenes, PHY-NeRF reaches 20.52 PSNR / 0.759 SSIM / 0.432 LPIPS on average, versus 19.87 PSNR for Aleth-NeRF.
- Ablation on low-light scenes shows removing the adaptive lighting particles drops average-scene PSNR to about 6 dB (chair 6.05, sofa 6.28).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PHY-NeRF instead of Aleth-NeRF | Novel-view exposure correction · Over-Exposure Scene (mean) | PSNR: 18.09 -> 21.38 | Helped |
| Use PHY-NeRF instead of Aleth-NeRF | Novel-view exposure correction · Under-Exposure Scene (mean) | PSNR: 19.87 -> 20.52 | Helped |
| Add Adaptive Lighting Particles α (ablation) | PHY-NeRF · Under-Exposure Scene, chair and sofa | PSNR: 6.05 -> 21.19; 6.28 -> 20.45 | Helped |
| Add image-level brightness consistency loss Lbc (ablation) | PHY-NeRF · Under-Exposure Scene, chair and sofa | PSNR: 11.97 -> 21.19; 15.46 -> 20.45 | Helped |
| Add spatial structure loss Lss (ablation) | PHY-NeRF · Under-Exposure Scene, chair and sofa | PSNR: 19.51 -> 21.19; 18.62 -> 20.45 | Helped |
| Add color constancy loss Lcc (ablation) | PHY-NeRF · Under-Exposure Scene, chair and sofa | PSNR: 20.95 -> 21.19; 19.49 -> 20.45 | Helped |
| Add illumination encoder Eθ (ablation) | PHY-NeRF · Under-Exposure Scene, chair and sofa | PSNR: 16.89 -> 21.19; 16.81 -> 20.45 | Helped |
Surprising findings
- Pre-processing images with exposure correction before NeRF training outperformed post-processing rendered NeRF outputs in the reported comparisons.
Limitations
- Gains on some individual scenes are small or negative (e.g., shrub underexposed PSNR 18.15 vs 18.24 for Aleth-NeRF).
- Relies on a separate reference-image set (LOL) for the illumination encoder.
- Evaluated only on five scenes from a single dataset.
Experimental setup
Datasets: Aleth-NeRF dataset (low-light, normal, overexposed multi-view scenes) LOL
Benchmarks: Aleth-NeRF multi-view dataset (5 overexposed and 5 underexposed scenes)
Metrics: PSNR SSIM LPIPS training time
Baselines: NeRF Aleth-NeRF HE + NeRF IAT + NeRF MSEC + NeRF RetiNexNet + NeRF Zero-DCE + NeRF SCI + NeRF
Builds on: NeRF Aleth-NeRF Zero-DCE MSGAN
Closest papers by meaning
- Aleth-NeRF: Illumination Adaptive NeRF with Concealing Field Assumption (AAAI 2024 · similarity 0.91)
- Enhancing Neural Radiance Fields with Adaptive Multi-Exposure Fusion: A Bilevel Optimization Approach for Novel View Synthesis (AAAI 2024 · similarity 0.91)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.90)
- A Pre-convolved Representation for Plug-and-Play Neural Illumination Fields (AAAI 2024 · similarity 0.89)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.89)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · 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.