Research map / Video / Video segmentation and tracking
DeblurSR: Event-Based Motion Deblurring under the Spiking Representation
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 DeblurSR, which models per-pixel latent intensity as a piecewise-linear spiking representation predicted from a blurry image and events, rendering sharp videos at arbitrary frame rates and extending to video super-resolution.
Problem
Event-based image-to-video motion deblurring: recovering a sharp, arbitrarily high frame-rate video from a single blurry frame and its events, where smooth parametric representations (polynomials) cannot capture intensity discontinuities.
Approach
Approximates each pixel's intensity over the exposure as a piecewise-linear function whose endpoint gaps model abrupt edge changes (spikes), with learned segment slopes, intercepts, and keypoint timings, plus an analytically solved integral normalization constant and a learned spatial kernel; coordinate embeddings allow sub-pixel querying for super-resolution.
Key results
- On REDS, DeblurSR reaches PSNR 26.725 and SSIM 0.859, a 12.3% MSE improvement over E-CIR (MSE 0.100 vs 0.114).
- On HQF, DeblurSR reaches PSNR 23.910 and SSIM 0.693, a 22.2% MSE improvement over E-CIR (MSE 0.161 vs 0.207).
- Training takes 72 hours on two V100 GPUs versus 100 hours on three V100 GPUs for E-CIR on REDS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace E-CIR with DeblurSR | Event-based motion deblurring · REDS | MSE: 0.114 -> 0.100 ± 0.001 | Helped |
| Replace E-CIR with DeblurSR | Event-based motion deblurring · REDS | PSNR: 25.531 -> 26.725 ± 0.001 | Helped |
| Replace E-CIR with DeblurSR | Event-based motion deblurring · REDS | SSIM: 0.819 -> 0.859 ± 0.001 | Helped |
| Replace E-CIR with DeblurSR | Event-based motion deblurring · HQF | MSE: 0.207 -> 0.161 ± 0.001 | Helped |
| Replace E-CIR with DeblurSR | Event-based motion deblurring · HQF | PSNR: 21.713 -> 23.910 ± 0.008 | Helped |
| Replace E-CIR with DeblurSR | Event-based motion deblurring · HQF | SSIM: 0.609 -> 0.693 ± 0.001 | Helped |
| Add automatic keypoint selection (ablation) | DeblurSR with heuristic keypoint selection · REDS | MSE: 0.107 ± 0.001 -> 0.103 ± 0.001 | Helped |
| Add automatic keypoint selection (ablation) | DeblurSR with heuristic keypoint selection · HQF | MSE: 0.236 ± 0.002 -> 0.229 ± 0.001 | Helped |
| Add integral normalization constant (ablation) | DeblurSR with automatic keypoint selection · REDS | MSE: 0.103 ± 0.001 -> 0.102 ± 0.001 | Helped |
| Add integral normalization constant (ablation) | DeblurSR with automatic keypoint selection · HQF | MSE: 0.229 ± 0.001 -> 0.163 ± 0.001 | Helped |
| Add kernel-based spatial modeling (ablation) | DeblurSR with automatic keypoint selection and integral normalization · REDS | MSE: 0.102 ± 0.001 -> 0.100 ± 0.001 | Helped |
| Add kernel-based spatial modeling (ablation) | DeblurSR with automatic keypoint selection and integral normalization · HQF | MSE: 0.163 ± 0.001 -> 0.161 ± 0.001 | Helped |
Surprising findings
- The integral normalization constant improves HQF MSE by 29.0% but only 1.0% on REDS, suggesting it matters most when data is limited and event noise is real.
Limitations
- Each pixel uses a fixed number of line segments (n=10); the paper notes varying segment counts per pixel would need a redesigned network.
- The HQF evaluation relies on only 9 training clips and 5 test clips, a small real-data benchmark.
Experimental setup
Datasets: REDS HQF
Benchmarks: REDS HQF
Metrics: MSE PSNR SSIM
Baselines: EDI eSL-Net E-CIR
Builds on: E-CIR Leaky-Integrate-and-Fire model UNet EDI
Closest papers by meaning
- SpikeReveal: Unlocking Temporal Sequences from Real Blurry Inputs with Spike Streams (NeurIPS 2024 · similarity 0.90)
- Spike-guided Motion Deblurring with Unknown Modal Spatiotemporal Alignment (CVPR 2024 · similarity 0.89)
- Spatio-Temporal Interactive Learning for Efficient Image Reconstruction of Spiking Cameras (NeurIPS 2024 · similarity 0.86)
- Motion-adaptive Transformer for Event-based Image Deblurring (AAAI 2025 · similarity 0.86)
- Event-Enhanced Blurry Video Super-Resolution (AAAI 2025 · similarity 0.86)
- Learning Scale-Aware Spatio-temporal Implicit Representation for Event-based Motion Deblurring (ICML 2024 · similarity 0.85)
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.