Research map / 3D vision / NeRF and radiance fields
Enhancing Neural Radiance Fields with Adaptive Multi-Exposure Fusion: A Bilevel Optimization Approach for Novel View Synthesis
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 AME-NeRF, an unsupervised NeRF framework that pairs a histogram-driven gamma exposure evaluator/simulator with HDR fusion and a bilevel optimization (sequential average method) to jointly correct multi-exposure inputs and synthesize novel views.
Problem
NeRF rendering degrades when training images have low or high exposure, since inconsistent luminance across views disrupts 3D consistency and 2D enhancement applied per view breaks geometry consistency.
Approach
Model exposure correction as a lower-level subproblem (gamma mapping chosen from histogram peak and skewness, fused via HDR weighting, refined by a U-Net) and NeRF rendering as the upper-level subproblem, solved jointly with a sequential average balancing of the two solutions, without paired supervision.
Key results
- On LOM low-exposure scenes, the method reaches PSNR 18.14 on 'bike' vs 18.02 for the best baseline (LIESMG+NeRF) and 6.36 for vanilla NeRF.
- On LOM high-exposure scenes, the method reaches PSNR 20.79 on 'bike' and 24.88 on 'chair', far above the best baselines (12.62 and 13.23 for DAConvUNet+NeRF).
- Ablation on LOM shows that adding both HDR fusion and the exposure evaluator (M4) gives the best averaged results (e.g., 19.47 PSNR on 'bike' vs 16.37 with HDR only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full AME-NeRF method instead of the strongest listed baseline | NeRF · LOM low exposure, bike | PSNR: 18.02 -> 18.14 | Helped |
| Use the full AME-NeRF method instead of the strongest listed baseline | NeRF · LOM low exposure, buu | PSNR: 16.21 -> 19.89 | Helped |
| Use the full AME-NeRF method instead of the strongest listed baseline | NeRF · LOM low exposure, chair | PSNR: 18.79 -> 17.05 | Hurt |
| Use the full AME-NeRF method instead of the strongest listed baseline | NeRF · LOM low exposure, shrub | PSNR: 17.63 -> 15.23 | Hurt |
| Use the full AME-NeRF method instead of the strongest listed baseline | NeRF · LOM low exposure, sofa | PSNR: 17.87 -> 17.93 | Helped |
| Use the full AME-NeRF method instead of the strongest listed baseline | NeRF · LOM high exposure | PSNR | Helped |
| Add HDR fusion and the exposure evaluator (ablation) | AME-NeRF · LOM average low- and high-exposure | PSNR: 5.99 -> 19.47; 6.53 -> 18.59; 6.08 -> 20.97; 6.08 -> 16.05; 6.27 -> 20.83 | Helped |
| Add HDR fusion (ablation) | AME-NeRF with exposure evaluator · LOM average low- and high-exposure | PSNR: 10.02 -> 19.47; 11.01 -> 18.59; 11.36 -> 20.97; 10.41 -> 16.05; 12.33 -> 20.83 | Helped |
| Add the exposure evaluator (ablation) | AME-NeRF with HDR fusion · LOM average low- and high-exposure | PSNR: 16.37 -> 19.47; 15.36 -> 18.59; 16.23 -> 20.97; 13.23 -> 16.05; 17.61 -> 20.83 | Helped |
| Replace random gamma values with the exposure evaluator's gamma selection (ablation) | AME-NeRF exposure correction · LOM | PSNR | Helped |
| Use HDR fusion instead of evaluator-selected gamma mapping alone (ablation) | AME-NeRF exposure correction · LOM | PSNR | Helped |
| Use multi-exposure inputs together with HDR images instead of HDR images alone (ablation) | AME-NeRF · LOM | content and structure consistency | Helped |
Surprising findings
- Multi-exposure input is necessary: HDR-only optimization degrades content and structure consistency, according to the authors' qualitative comparison.
Limitations
- Evaluation limited to five LOM scenes plus LOL single-image enhancement; overexposed LOM is self-constructed.
- Bilevel balancing via a fixed alpha-weighted combination and hand-chosen hyperparameters (a, b, c, d, alpha) lack broader sensitivity analysis.
- Results are reported only with PSNR/SSIM/LPIPS and visual comparisons, with no statistical variance.
Experimental setup
Datasets: LOM LOL
Benchmarks: LOM LOL
Metrics: PSNR SSIM LPIPS
Baselines: NeRF HE + NeRF IAT + NeRF LIESMG + NeRF LIME + NeRF SCI + NeRF DAConvUNet + NeRF LRS + NeRF AHE + NeRF CLAHE + NeRF
Builds on: NeRF Sequential Average Method U-Net VGG16 perceptual loss
Closest papers by meaning
- Physical-aware Neural Radiance Fields for Efficient Exposure Correction (AAAI 2025 · similarity 0.91)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.91)
- Few-Shot Neural Radiance Fields under Unconstrained Illumination (AAAI 2024 · similarity 0.90)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.90)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.90)
- 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.