Research map / Image generation / Image restoration
Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach
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 the SDE dataset of over 30K spatio-temporally aligned real-world image-event pairs captured with a robotic arm, and EvLight, an event-guided low-light image enhancement network using SNR-guided regional feature selection and holistic-regional fusion.
Problem
Event-guided low-light image enhancement is limited by the lack of large-scale, real-world, spatio-temporally aligned image-event datasets, and existing fusion methods ignore region-wise noise discrepancies between images and events.
Approach
Capture aligned low/normal-light image-event pairs with a UR5 robotic arm following non-linear trajectories plus a matching temporal alignment strategy, and fuse modalities holistically (attention UNet) and regionally, using an SNR map to select image features in high-SNR regions and event features in low-SNR regions.
Key results
- On the SDE dataset, EvLight reaches 22.44 dB PSNR on SDE-in and 23.21 dB on SDE-out, exceeding the best baseline by 0.65 dB and 0.29 dB respectively.
- On SDSD-out, EvLight improves PSNR* over the best baseline by more than 0.59 dB and beats Retinexformer by 2.62 dB PSNR (as stated in the abstract).
- Ablation shows SNR-guided regional feature selection adds 0.86 dB PSNR over the base model on SDE-in.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EvLight instead of the strongest baseline | Low-light image enhancement · SDE-in | PSNR: 21.79 -> 22.44 | Helped |
| use EvLight instead of the strongest baseline | Low-light image enhancement · SDE-out | PSNR: 22.92 -> 23.21 | Helped |
| use EvLight instead of the strongest baseline | Low-light image enhancement · SDSD-in | PSNR: 27.58 -> 28.52 | Helped |
| use EvLight instead of the strongest baseline | Low-light image enhancement · SDSD-out | PSNR: 26.08 -> 26.67 | Helped |
| add events to the Base model (ablation) | Base model · SDE-in | PSNR: +0.23 dB | Helped |
| add SNR-guided regional feature selection instead of no regional selection (ablation) | Base model · SDE-in | PSNR: 21.58 -> 22.44; 0.86 dB increase | Helped |
| use regional feature selection with an all-ones matrix instead of no selection (ablation) | Base model · SDE-in | PSNR: 21.58 -> 21.86; 0.28 dB increase | Helped |
| add image-regional feature selection (IRFS) (ablation) | EvLight · SDE-in | PSNR: 21.58 -> 21.92; 0.34 dB increase | Helped |
| add event-regional feature selection (ERFS) (ablation) | EvLight · SDE-in | PSNR: 21.58 -> 22.18; 0.60 dB increase | Helped |
| add both IRFS and ERFS (ablation) | Base model · SDE-in | PSNR: 21.58 -> 22.44; 0.86 dB increase | Helped |
Surprising findings
- Event-fusion methods ELIE and Liu et al. beat frame-only methods on SDSD-in but overfit on SDSD-out, showing a large PSNR vs PSNR* gap.
Limitations
- DAVIS346 hardware causes partial chromatic aberrations and moiré patterns in RGB images
- Image and event triggering is not synchronized, requiring repeated captures and labor-intensive collection
- SNR map also assigns low values to edge-rich regions, which the method must compensate for with events
Experimental setup
Datasets: SDE (new) SDSD
Benchmarks: SDE dataset (SDE-in, SDE-out) SDSD (in/out) CED MVSEC
Metrics: PSNR PSNR* SSIM
Baselines: E2VID+ SNR-Net Uformer LLFlow-L-SKF Retinexformer ELIE eSL-Net Liu et al. (AAAI'23)
Builds on: Retinexformer SNR-Net ELIE v2e E2VID
Closest papers by meaning
- LED: A Large-scale Real-world Paired Dataset for Event Camera Denoising (CVPR 2024 · similarity 0.90)
- Seeing Motion at Nighttime with an Event Camera (CVPR 2024 · similarity 0.88)
- Event-assisted Low-Light Video Object Segmentation (CVPR 2024 · similarity 0.87)
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.87)
- Event Fields: Capturing Light Fields at High Speed, Resolution, and Dynamic Range (CVPR 2025 · similarity 0.87)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · 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.