Research map / Video / Video segmentation and tracking
Motion Blur Decomposition with Cross-shutter Guidance
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 dual Blur-RS (cross-shutter) input for motion blur decomposition, using a rolling-shutter view's row-wise temporal ordering to disambiguate a blurry image's latent sharp frames, with a triaxial capture rig and the RealBR dataset.
Problem
Decomposing a single motion-blurred image into an ordered sharp video is highly ambiguous because averaging destroys temporal order, and existing methods need neighboring blurry frames or human annotation for motion disambiguation.
Approach
Pair the blurry image with a rolling-shutter image whose scanline-wise exposure delay encodes temporal order, and process the two views in mutually boosting branches (shutter alignment and aggregation plus temporal positional encoding) to decompose the blur into sharp frames.
Key results
- On RealBR (x9 latent frames), the proposed model reaches 31.15 dB PSNR / 0.9120 SSIM, versus 24.83 dB for RIFE on blur-only input and 30.84 dB for IFED on blur-RS input.
- Using a blur-RS dual view improves RIFE by at least 5.6 dB PSNR over blur-only consecutive-frame variants.
- Under misaligned RS views (max shift 8 px), PSNR drops only slightly, to 30.87 dB at x9 frames, indicating robustness to misalignment.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add temporal positional encoding to the RS branch (ablation) | Ours · RealBR | PSNR: 31.06 -> 31.15 | Helped |
| Add shutter alignment and aggregation (SAA) (ablation) | Ours · RealBR | PSNR: 27.96 -> 31.15 | Helped |
| Use dual motion interpretation branches instead of a single branch (ablation) | Ours · RealBR | PSNR: 30.33 -> 31.15 | Helped |
| Use Blur-RS input instead of neighboring blur frames | RIFE · RealBR ×3 | PSNR: 24.60 -> 30.26 | Helped |
| Use Blur-RS input instead of neighboring blur frames | IFED · RealBR ×3 | PSNR: 24.45 -> 30.46 | Helped |
| Add Blur view to RS temporal super-resolution input | RIFE · RealBR ×5 | PSNR: 24.14 -> 30.53 | Helped |
Surprising findings
- Misaligning the RS view by up to 8 pixels causes little performance loss, suggesting the RS view serves mainly as a motion guide rather than requiring precise alignment.
Limitations
- Requires a beamsplitter triaxial rig, which is impractical for compact mobile devices.
- Requires an extra RS view, increasing device cost compared with single-image or neighboring-frame methods.
Experimental setup
Datasets: RealBR synthetic Blur-RS data
Benchmarks: RealBR synthetic Blur-RS data
Metrics: PSNR SSIM LPIPS running time parameters FLOPs
Baselines: LEVS AfB BiT RIFE IFED DeMFI RSSR CVR PMB EBFI EvUnroll
Builds on: MIMO-UNet RIFE IFED BiT AfB
Closest papers by meaning
- Context-Aware Video Reconstruction for Rolling Shutter Cameras (CVPR 2022 · similarity 0.88)
- Spatio-Temporal Difference Guided Motion Deblurring with the Complementary Vision Sensor (CVPR 2026 · similarity 0.87)
- Joint Video Multi-Frame Interpolation and Deblurring Under Unknown Exposure Time (CVPR 2023 · similarity 0.87)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.86)
- Motion Deblurring via Spatial-Temporal Collaboration of Frames and Events (AAAI 2024 · similarity 0.85)
- Motion-adaptive Transformer for Event-based Image Deblurring (AAAI 2025 · 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.