Research map / Image generation / Image restoration
Single Image Rolling Shutter Removal with Diffusion Models
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 RS-Diffusion, a diffusion-based single-image rolling shutter correction method that predicts an IGF (gyro-derived motion field) with a patch-attention module, plus the RS-Real dataset labeled via IMU gyroscope data.
Problem
Single-frame rolling shutter (RS) correction, which is harder than multi-frame correction because it lacks temporal cues and existing RS datasets fail to combine accurate labels with realistic content.
Approach
Treats RS correction as an image-to-motion problem: a conditional diffusion model (CFG + DDIM) generates an intra gyro field (motion field) from the RS image, with a patch-attention block exploiting row-wise RS motion correlation; ground-truth motion fields come from synchronized gyroscope rotations converted to homographies.
Key results
- On the Building benchmark RS-Diffusion reaches 34.92 dB PSNR, 0.79 SSIM and 0.90 EPE, beating Yan et al. (33.34 dB, 0.75, 1.25).
- On RS-Homo it reports 36.60 dB PSNR vs 26.15 dB for Yan et al.; on RS-Real it reports 22.02 dB vs 18.48 dB.
- Predicting IGF rather than GS images improves PSNR from 16.73 to 20.78 dB in the ablation; adding intra and inter patch attention raises it to 22.02 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| predict IGF from RS images instead of predicting GS images directly (ablation) | RS-Diffusion · RS-Real | PSNR (dB): 16.73 -> 20.78 | Helped |
| predict IGF from RS images instead of predicting GS images directly (ablation) | RS-Diffusion · RS-Real | SSIM: 0.47 -> 0.63 | Helped |
| add intra-patch attention (ablation) | RS-Diffusion · RS-Real | PSNR (dB): 20.78 -> 21.10 | Helped |
| add intra-patch attention (ablation) | RS-Diffusion · RS-Real | SSIM: 0.63 -> 0.65 | Helped |
| add intra-patch attention (ablation) | RS-Diffusion · RS-Real | EPE: 2.62 -> 2.54 | Helped |
| add inter-patch attention to intra-patch attention (ablation) | RS-Diffusion · RS-Real | PSNR (dB): 21.10 -> 22.02 | Helped |
| add inter-patch attention to intra-patch attention (ablation) | RS-Diffusion · RS-Real | SSIM: 0.65 -> 0.69 | Helped |
| add inter-patch attention to intra-patch attention (ablation) | RS-Diffusion · RS-Real | EPE: 2.54 -> 2.12 | Helped |
| replace Yan et al. with RS-Diffusion | single-image rolling shutter correction · Building | PSNR (dB): 33.34 -> 34.92 | Helped |
| replace Yan et al. with RS-Diffusion | single-image rolling shutter correction · RS-Homo | PSNR (dB): 26.15 -> 36.60 | Helped |
| replace Yan et al. with RS-Diffusion | single-image rolling shutter correction · RS-Real | PSNR (dB): 18.48 -> 22.02 | Helped |
Limitations
- Gyroscope-based labels model rotation only and ignore translation and depth effects.
- Only a small number of baselines are compared on RS-Real, since Yan et al. is the only open-source competitor retrained on it.
Experimental setup
Datasets: Building RS-Homo RS-Real
Benchmarks: Building RS-Homo RS-Real
Metrics: PSNR SSIM EPE
Baselines: Rengarajan et al. Purkait et al. Grundmann et al. Kandula et al. Yan et al. deep homography mixture model
Builds on: DDPM DDIM classifier-free guidance Yan et al. deep homography mixture model
Closest papers by meaning
- Learning Adaptive Warping for Real-World Rolling Shutter Correction (CVPR 2022 · similarity 0.89)
- Rolling Shutter Correction with Intermediate Distortion Flow Estimation (CVPR 2024 · similarity 0.87)
- USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields (ICLR 2024 · similarity 0.85)
- Order-One Rolling Shutter Cameras (CVPR 2025 · similarity 0.84)
- Motion Blur Decomposition with Cross-shutter Guidance (CVPR 2024 · similarity 0.84)
- EVS-assisted Joint Deblurring Rolling-Shutter Correction and Video Frame Interpolation through Sensor Inverse Modeling (CVPR 2024 · similarity 0.84)
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.