Research map / 3D vision / NeRF and radiance fields
EvHDR-NeRF: Building High Dynamic Range Radiance Fields with Single Exposure Images and Events
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.
Recovers an HDR NeRF from single-exposure LDR views plus event streams by modeling events, radiance, CRF and exposure time, with two MLPs and an event-based loss.
Problem
HDR radiance field reconstruction of 3D scenes from single-exposure LDR views is ill-posed; existing NeRF HDR methods need multi-exposure inputs and event-guided HDR methods ignore CRF and exposure time and lack multi-view consistency.
Approach
Models the physical relation between events, radiance, and LDR frames via a learnable CRF (MLP G) and exposure integration; inter-frame events supervise an HDR radiance field (MLP F) while intra-exposure events help deblur and constrain LDR color.
Key results
- On the Over-Exposed synthetic set, EvHDR-NeRF reaches 28.95 PSNR / 0.9162 SSIM / 0.0586 LPIPS vs 24.15 PSNR for Huang et al. (HDR-NeRF).
- On the Blur synthetic set, it reaches 28.32 PSNR vs 15.73 for Huang et al. and 15.86 for Yang et al.
- Novel-view synthesis on the Over-Exposed set achieves 29.42 PSNR vs 24.03 for Huang et al.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EvHDR-NeRF instead of the strongest listed baseline | HDR reconstruction · Over-Exposed (input view) | PSNR: 24.70 -> 28.95 | Helped |
| use EvHDR-NeRF instead of the strongest listed baseline | HDR reconstruction · Blur (input view) | PSNR: 15.86 -> 28.3245 | Helped |
| use EvHDR-NeRF instead of the strongest listed baseline | HDR reconstruction · Under-Exposed (input view) | PSNR: 17.9680 -> 28.01 | Helped |
| use EvHDR-NeRF instead of Huang et al. | HDR reconstruction · Over-Exposed (novel view) | PSNR: 24.0275 -> 29.4243 | Helped |
| use EvHDR-NeRF instead of Huang et al. | HDR reconstruction · Blur (novel view) | PSNR: 16.0738 -> 28.8037 | Helped |
| use EvHDR-NeRF instead of Huang et al. | HDR reconstruction · Under-Exposed (novel view) | PSNR: 17.8692 -> 27.7480 | Helped |
| set the HDR-to-LDR gamma value to 1/5 instead of 1/2.2 (ablation) | EvHDR-NeRF · Synthetic dog scene | CRF estimation | No clear effect |
Surprising findings
- Huang et al. HDR-NeRF struggles to learn an accurate CRF unless the input exposure range is wide; the proposed method approximates it with a single exposure.
Limitations
- Less accurate CRF estimation at very short exposure times because events give only incremental information.
- Real-world evaluation limited to three scenes captured with a single DAVIS 346C sensor.
- Evaluated mainly on synthetic data built by augmenting HDR-NeRF scenes with simulated events.
Experimental setup
Datasets: HDR-NeRF synthetic scenes (Huang et al. 2022) augmented with blur, low-light and events real-world DAVIS 346C captures (3 scenes)
Benchmarks: Over-Exposed synthetic Blur synthetic Under-Exposed synthetic real-world DAVIS 346C captures
Metrics: PSNR SSIM LPIPS
Baselines: Huang et al. (HDR-NeRF) Yang et al. (event-guided HDR) Wu et al. (dynamic HDR NeRF) Rudnev et al. Low and Lee Klenk et al. Huang et al. + X
Builds on: NeRF HDR-NeRF event-based image reconstruction (Yang et al. 2023) NeRF-based event supervision (Rudnev et al. 2023)
Closest papers by meaning
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.89)
- Pano-NeRF: Synthesizing High Dynamic Range Novel Views with Geometry from Sparse Low Dynamic Range Panoramic Images (AAAI 2024 · similarity 0.89)
- EvHDR-GS: Event-guided HDR Video Reconstruction with 3D Gaussian Splatting (AAAI 2025 · similarity 0.88)
- Color Event Enhanced Single-Exposure HDR Imaging (AAAI 2024 · similarity 0.88)
- Single Shot HDR Recovery via a Video Diffusion Prior (NeurIPS 2026 · similarity 0.88)
- HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting (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.