Research map / Video / Video segmentation and tracking
EvSTVSR: Event Guided Space-Time Video Super-Resolution
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 EvSTVSR, an event-guided space-time video super-resolution method that uses only two adjacent LR frames plus intervening events, with event-guided optical flow alignment and coordinate-based (SIREN) fusion upsampling.
Problem
Space-time video super-resolution struggles with large, nonlinear motion and varying illumination when only RGB frames are available, because inter-frame information is missing and multi-frame alignment accumulates errors.
Approach
Uses two RGB frames plus the events between them, estimating optical flow from both images and events to warp frames and features to the target time, then fuses warping and synthesis paths with a positional-coordinate (SIREN-based) implicit upsampling network and a transformer decoder.
Key results
- On Adobe240 x8 VFI / x4 VSR, achieves 31.61 dB PSNR Adobe-Average versus 31.11 dB for event-based HR-INR and 29.82 dB for RGB-based MoTIF.
- On the CED VSR x4 benchmark, reaches 33.83 dB PSNR, above EvTexture (33.68 dB) and HR-INR (32.15 dB).
- On BS-ERGB VFI at Time x2, reaches 29.73 dB PSNR versus 29.66 dB for HR-INR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EvSTVSR instead of HR-INR | STVSR · GoPro-Center | PSNR: 31.97 -> 32.50 | Helped |
| use EvSTVSR instead of HR-INR | STVSR · GoPro-Average | PSNR: 32.13 -> 32.23 | Helped |
| use EvSTVSR instead of HR-INR | STVSR · Adobe-Center | PSNR: 31.26 -> 31.79 | Helped |
| use EvSTVSR instead of HR-INR | STVSR · Adobe-Average | PSNR: 31.11 -> 31.61 | Helped |
| use EvSTVSR instead of HR-INR | VFI · BS-ERGB 1skip | PSNR: 29.66 -> 29.73 | Helped |
| use EvSTVSR instead of HR-INR | VFI · BS-ERGB 3skips | PSNR: 28.59 -> 28.64 | Helped |
| use EvSTVSR instead of EvTexture+ | VSR · CED Space ×4 | PSNR: 33.71 -> 33.83 | Helped |
| use EvSTVSR instead of HR-INR | VSR · CED Space ×2 | PSNR: 42.01 -> 42.47 | Helped |
| add explicit optical-flow alignment (ablation) | EvSTVSR · Adobe-Center | PSNR: 30.92 -> 31.79 | Helped |
| add explicit optical-flow alignment (ablation) | EvSTVSR · Adobe-Center | SSIM: 0.9123 -> 0.9200 | Helped |
| add explicit optical-flow alignment (ablation) | EvSTVSR · Adobe-Average | PSNR: 30.88 -> 31.61 | Helped |
| add explicit optical-flow alignment (ablation) | EvSTVSR · Adobe-Average | SSIM: 0.9121 -> 0.9194 | Helped |
Surprising findings
- Using fewer input frames (two or three) outperforms using five frames, which the authors attribute to alignment difficulty under complex motion.
- Removing optical flow alignment reduces performance only modestly on Adobe240 (31.61 to 30.88 dB average), yet the paper still reports gains over multi-frame HR-INR with just two inputs.
Limitations
- Training process is time-consuming.
- Generalization to real event data remains a challenge.
- Slightly lower SSIM than HR-INR on some metrics because it uses only two image inputs.
Experimental setup
Datasets: Adobe240 GoPro BS-ERGB CED
Benchmarks: Adobe240 GoPro BS-ERGB CED
Metrics: PSNR SSIM
Baselines: SuperSloMo QVI DAIN EDVR BasicVSR Zooming SlowMo TMNet VideoINR MoTIF HR-INR TimeLens TimeLens++
Builds on: VideoINR CBMNet LIIF SIREN TimeLens
Closest papers by meaning
- Time Without Time: Pseudo-Temporal Representation for Space-Time Super-Resolution (CVPR 2026 · similarity 0.90)
- EvEnhancer: Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events (CVPR 2025 · similarity 0.89)
- Event-Enhanced Blurry Video Super-Resolution (AAAI 2025 · similarity 0.88)
- EvTexture: Event-driven Texture Enhancement for Video Super-Resolution (ICML 2024 · similarity 0.88)
- Event-based Video Super-Resolution via State Space Models (CVPR 2025 · similarity 0.87)
- BF-STVSR: B-Splines and Fourier---Best Friends for High Fidelity Spatial-Temporal Video Super-Resolution (CVPR 2025 · similarity 0.86)
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.