Research map / Video / Video segmentation and tracking
EvTexture: Event-driven Texture Enhancement for 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 EvTexture, the first video super-resolution method that uses event signals for texture enhancement via a parallel texture branch with an iterative GRU-based texture updater over event voxel bins.
Problem
Video super-resolution struggles to restore high-frequency texture regions from LR frames; prior event-based VSR methods use events mainly to improve motion/flow estimation rather than texture.
Approach
Add a texture enhancement branch alongside the flow-based motion branch in BasicVSR, where an Iterative Texture Enhancement module uses a ConvGRU updater to progressively inject high-frequency details from event voxel bins (one iteration per bin) into the propagation feature.
Key results
- On Vid4 (4x), EvTexture reaches 29.51 dB average PSNR, and EvTexture+ 29.78 dB, versus 24.84 dB for EGVSR (up to +4.67 dB gain over event-based methods).
- On CED (real-world), EvTexture improves over BasicVSR by +0.95 dB at 2x (40.52 vs 39.57 dB PSNR) and on 4x reaches 33.68 dB vs EBVSR 33.42 dB.
- On REDS4 4x, EvTexture achieves 32.79 dB vs BasicVSR 31.42 dB, and ablation shows the texture branch alone gives 29.22 dB on Vid4 versus 27.44 for the motion-only branch.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add an event-driven texture enhancement branch to the motion branch (ablation) | EvTexture · Vid4 | PSNR: +2.07dB | Helped |
| Combine the motion learning and texture enhancement branches (ablation) | EvTexture · REDS4 | PSNR | Helped |
| Replace Conv layers with a ConvGRU texture updater (ablation) | ITE module · Vid4 | PSNR: 29.174 -> 29.507 | Helped |
| Use iterative updates instead of direct UNet extraction from the whole voxel grid (ablation) | ITE module · Vid4 | PSNR: 0.42dB | Helped |
| Add residual learning to the texture updater (ablation) | ITE module · Vid4 | PSNR: 0.35dB | Helped |
| Use 5 iterations instead of 3 (ablation) | ITE module · Vid4 | PSNR: 29.463 -> 29.507 | Helped |
| Use 5 iterations instead of 8 (ablation) | ITE module · Vid4 | PSNR: 29.324 -> 29.507 | Helped |
| Add event-driven motion enhancement to EvTexture | EvTexture · Vid4, REDS4, Vimeo-90K-T, and CED | PSNR | Helped |
| Add event signals to texture restoration rather than using the RGB-based VRT baseline | 4× VSR · Vid4 | PSNR: +1.63dB | Helped |
| Use EvTexture instead of EBVSR | 4× VSR · REDS4 | PSNR: +1.32 | Helped |
Surprising findings
- Increasing the number of iterations beyond 5 (matching voxel bin count) degrades performance, attributed to sparse, noisy voxel bins at 8 bins.
Limitations
- Evaluation on synthetic events for Vid4, REDS, and Vimeo-90K (simulated with ESIM), with only CED providing real events.
- Requires equal spatial resolution between frames and events; asymmetric resolution is left to future work.
- Texture magnitude difficulty split for Vimeo-90K-T uses ad hoc percentage thresholds.
Experimental setup
Datasets: Vimeo-90K REDS Vid4 REDS4 Vimeo-90K-T CED
Benchmarks: Vid4 REDS4 Vimeo-90K-T CED
Metrics: PSNR SSIM LPIPS runtime parameters
Baselines: EDVR BasicVSR IconVSR RVRT BasicVSR++ VRT EGVSR EBVSR E-VSR RBPN TDAN DUF
Builds on: BasicVSR RAFT SpyNet EBVSR
Closest papers by meaning
- EvSTVSR: Event Guided Space-Time Video Super-Resolution (AAAI 2025 · similarity 0.88)
- Event-Enhanced Blurry Video Super-Resolution (AAAI 2025 · similarity 0.86)
- EvEnhancer: Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events (CVPR 2025 · similarity 0.84)
- VideoGigaGAN: Towards Detail-rich Video Super-Resolution (CVPR 2025 · similarity 0.84)
- TTA-EVF: Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation (CVPR 2024 · similarity 0.83)
- Frequency-aware Event-based Video Deblurring for Real-World Motion Blur (CVPR 2024 · 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.