Research map / Image generation / Image restoration
Color Event Enhanced Single-Exposure HDR Imaging
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.
Introduces color event cameras into single-exposure HDR imaging, fusing an LDR image with color event voxels through an exposure-aware transformer that propagates color hints into over/under-exposed regions.
Problem
Single-exposure HDR reconstruction from one LDR image cannot recover color in severely over- or under-exposed regions because the information is lost in the LDR capture.
Approach
Adds color event streams (RGBG-filtered event camera) as auxiliary input and uses an exposure-aware transformer with a learned hint-selected mask to propagate color hints from normal-exposed regions and events into missing regions, plus a region-based discriminator on over/under-exposed areas.
Key results
- On HDR-SYNTH, the full model reaches HDR-VDP-3 7.39 ± 1.16, versus 6.92 for SingleHDR (the prior best).
- On HDR-REAL, the method reaches 6.04 ± 1.68 with SYNTH, REAL training, versus 5.93 for SingleHDR.
- On the real-world HDR-CE dataset, the method reaches 7.56 ± 1.14 versus 7.35 for SingleHDR fine-tuned with SYNTH, CE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add color events to the single-exposure HDR model (ablation) | Ours · HDR-SYNTH | HDR-VDP-3: 6.98 ± 1.51 -> 7.14 ± 1.33 | Helped |
| add the exposure-aware transformer (ablation) | Ours · HDR-SYNTH | HDR-VDP-3: 7.14 ± 1.33 -> 7.39 ± 1.16 | Helped |
| add the hint-selected mask to EaT (ablation) | Ours · HDR-SYNTH | HDR-VDP-3: 7.29 ± 1.20 -> 7.39 ± 1.16 | Helped |
| add image-event fusion blocks instead of concatenating event and image features (ablation) | Ours · HDR-SYNTH | HDR-VDP-3: 7.34 ± 1.14 -> 7.39 ± 1.16 | Helped |
| merge fused features in the decoding stage rather than the encoding stage (ablation) | Ours · HDR-SYNTH | HDR-VDP-3: 7.29 ± 1.21 -> 7.39 ± 1.16 | Helped |
| add the region-based discriminator (ablation) | Ours · HDR-SYNTH | HDR-VDP-3: 7.32 ± 1.22 -> 7.39 ± 1.16 | Helped |
| use the full color-event method instead of SingleHDR | single-exposure HDR imaging · HDR-SYNTH | HDR-VDP-3: 6.92 ± 1.67 -> 7.39± 1.16 | Helped |
| use the full color-event method instead of SingleHDR | single-exposure HDR imaging · HDR-REAL | HDR-VDP-3: 5.93 ± 2.01 -> 6.04 ± 1.68 | Helped |
| use the full color-event method instead of SingleHDR | single-exposure HDR imaging · HDR-CE | HDR-VDP-3: 7.35 ± 1.32 -> 7.56 ± 1.14 | Helped |
Limitations
- Improvements over the strongest baseline are modest (e.g., 0.11 on HDR-REAL).
- Some baselines were pre-trained models rather than retrained, so comparisons may be uneven.
- Evaluation relies only on HDR-VDP-3 with no reported color-specific metric.
Experimental setup
Datasets: HDR-SYNTH HDR-REAL HDR-CE
Benchmarks: HDR-SYNTH HDR-REAL HDR-CE
Metrics: HDR-VDP-3
Baselines: HDRCNN DrTMO ExpandNet SingleHDR Diff KUNET AEO HPEO KOEO MEO
Builds on: SingleHDR ESIM ConvLSTM Nazeri et al. region-based discriminator
Closest papers by meaning
- Single Shot HDR Recovery via a Video Diffusion Prior (NeurIPS 2026 · similarity 0.88)
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.88)
- EvHDR-NeRF: Building High Dynamic Range Radiance Fields with Single Exposure Images and Events (AAAI 2025 · similarity 0.88)
- Self-Supervised High Dynamic Range Imaging with Multi-Exposure Images in Dynamic Scenes (ICLR 2024 · similarity 0.87)
- EvHDR-GS: Event-guided HDR Video Reconstruction with 3D Gaussian Splatting (AAAI 2025 · similarity 0.86)
- UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion (CVPR 2025 · similarity 0.86)
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.