Research map / 3D vision / NeRF and radiance fields
Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW 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.
Bright-NeRF learns an unsupervised, ISP-free normal-light radiance field from multi-view low-light raw images using a learned von Kries-style chromatic adaptation field, a smoothness loss, and exposure scaling, and introduces the LMRAW dataset.
Problem
Novel view synthesis from multi-view low-light images is degraded by noise, low brightness and color distortion, and existing 2D enhancement and NeRF methods either overfit to a fixed ISP or depend on a pre-calibrated camera ISP.
Approach
Learn a view-dependent diagonal sensor-response (von Kries chromatic adaptation) field from NeRF features, supervised by a chromatic adaptation loss, and rely on multi-view aggregation for denoising and raw linear scaling for exposure, trained only on low-light raw images without normal-light supervision.
Key results
- Bright-NeRF achieves 22.64 PSNR, 0.874 SSIM, 0.476 LPIPS averaged over nine LMRAW scenes, versus 18.60 PSNR for the best 2D baseline DNF and 16.64 PSNR for LLNeRF.
- Ablation shows removing the color restoration field and chromatic adaptation loss drops PSNR from 22.61 to 21.37 (with smoothness loss also removed).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Bright-NeRF instead of LLNeRF | LMRAW novel-view synthesis · LMRAW (nine-scene mean) | PSNR: 16.64 -> 22.64 | Helped |
| use Bright-NeRF instead of LLNeRF | LMRAW novel-view synthesis · LMRAW (nine-scene mean) | SSIM: 0.811 -> 0.874 | Helped |
| use Bright-NeRF instead of LLNeRF | LMRAW novel-view synthesis · LMRAW (nine-scene mean) | LPIPS: 0.494 -> 0.476 | Helped |
| use Bright-NeRF instead of LLNeRF | LMRAW novel-view synthesis · LMRAW (nine-scene mean) | LIQE: 1.332 -> 1.659 | Helped |
| use Bright-NeRF instead of LLNeRF | LMRAW enhancement with sRGB images · LMRAW (nine-scene mean, sRGB images) | PSNR: 16.64 -> 21.49 | Helped |
| remove color restoration and chromatic adaptation loss (ablation) | Bright-NeRF · LMRAW ablation | PSNR: 22.64 -> 21.37 | Hurt |
| remove color restoration and chromatic adaptation loss (ablation) | Bright-NeRF · LMRAW ablation | LPIPS: 0.476 -> 0.737 | Hurt |
| remove local smoothness loss (ablation) | Bright-NeRF · LMRAW ablation | PSNR: 22.64 -> 22.61 | Hurt |
| remove local smoothness loss (ablation) | Bright-NeRF · LMRAW ablation | LIQE: 1.659 -> 1.614 | Hurt |
Surprising findings
- RawNeRF scores very low PSNR (8.95) on LMRAW despite high LIQE, indicating its generic-ISP output is washed out.
- LLNeRF's color bias grows as scenes get darker, while Bright-NeRF stays stable across exposure ratios.
Limitations
- Evaluated on only nine scenes in a single controlled dark-room dataset.
- Uses a learned sensor-response model with hand-set loss weights (1.0, 0.1, 0.1), with no sensitivity analysis reported.
Experimental setup
Datasets: LMRAW
Benchmarks: LMRAW
Metrics: PSNR SSIM LPIPS LIQE
Baselines: SID RED ABF DNF RawNeRF DNF+NeRF LLNeRF
Builds on: NeRF RawNeRF LLNeRF von Kries chromatic adaptation
Closest papers by meaning
- LuSh-NeRF: Lighting up and Sharpening NeRFs for Low-light Scenes (NeurIPS 2024 · similarity 0.91)
- Few-Shot Neural Radiance Fields under Unconstrained Illumination (AAAI 2024 · similarity 0.90)
- IReNe: Instant Recoloring of Neural Radiance Fields (CVPR 2024 · similarity 0.90)
- Enhancing Neural Radiance Fields with Adaptive Multi-Exposure Fusion: A Bilevel Optimization Approach for Novel View Synthesis (AAAI 2024 · similarity 0.90)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.90)
- Aleth-NeRF: Illumination Adaptive NeRF with Concealing Field Assumption (AAAI 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.