Research map / Video / Video segmentation and tracking
EvDiG: Event-guided Direct and Global Components Separation
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.
Separates direct and global illumination components from a single RGB frame plus event-camera recordings of a sweeping line-occluder shadow, using event accumulation followed by two networks (EvSepNet, ImColorNet).
Problem
Direct-global illumination separation from RGB images is ill-posed for single frames, while multi-frame active-illumination methods need many captures with a static scene and long capture times.
Approach
Uses an event camera to record rapid illumination changes from a line occluder's shadow sweep; event accumulation yields coarse max/min brightness and thus coarse direct/global components, which are refined by a U-Net with spatial attention (SCB) and colorized by a transposed-attention color correction network (ImColorNet).
Key results
- On the self-collected dataset, EvDiG reaches 30.01 PSNR / 0.883 SSIM / 0.077 LPIPS on direct components, versus 28.51 PSNR for the best baseline SF-pattern-deep.
- Capture time is 33 ms (single-frame equivalent) versus 833 ms for 25-frame multi-frame pattern methods, about 20x more efficient.
- On global components EvDiG gets 31.66 PSNR, below SF-pattern-deep's 32.96 PSNR, but has better LPIPS (0.117 vs 0.228).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use event-guided EvDiG instead of SF-pattern-deep | Direct-component separation · Indoor dataset | PSNR: 28.51 -> 30.01 | Helped |
| Use event-guided EvDiG instead of SF-pattern-deep | Global-component separation · Indoor dataset | PSNR: 32.96 -> 31.66 | Hurt |
| Use event-guided EvDiG instead of SF-pattern-deep | Direct-component separation · Indoor dataset | SSIM: 0.834 -> 0.883 | Helped |
| Use event-guided EvDiG instead of SF-pattern-deep | Global-component separation · Indoor dataset | SSIM: 0.867 -> 0.846 | Hurt |
| Use event-guided EvDiG instead of SF-pattern-deep | Direct-component separation · Indoor dataset | LPIPS: 0.208 -> 0.077 | Helped |
| Use event-guided EvDiG instead of SF-pattern-deep | Global-component separation · Indoor dataset | LPIPS: 0.228 -> 0.117 | Helped |
Surprising findings
- On global components, EvDiG has lower PSNR and SSIM than SF-pattern-deep (31.66 vs 32.96 PSNR), though it leads on LPIPS.
Limitations
- Reference ground truth comes from Nayar et al.'s method, not true ground truth.
- Scene motion events are not modeled; method works only when scene motion is small relative to shadow motion.
Experimental setup
Datasets: self-collected RGB-event indoor dataset ImageNet
Benchmarks: self-collected indoor dataset outdoor scenes (qualitative)
Metrics: PSNR SSIM LPIPS
Baselines: MF-shadow-classic SF-pattern-classic SF-scene-deep SF-pattern-deep
Builds on: Nayar et al. high-frequency illumination separation Restormer event-based video reconstruction multi-scale architecture
Closest papers by meaning
- EvGGS: A Collaborative Learning Framework for Event-based Generalizable Gaussian Splatting (ICML 2024 · similarity 0.86)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.85)
- Event-Driven Dynamic Scene Depth Completion (NeurIPS 2025 · similarity 0.85)
- Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach (CVPR 2024 · similarity 0.84)
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.84)
- EvFocus: Learning to Reconstruct Sharp Images from Out-of-Focus Event Streams (ICML 2025 · similarity 0.83)
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.