Research map / Video / Video segmentation and tracking
EvEnhancer: Empowering Effectiveness, Efficiency and Generalizability for Continuous Space-Time Video Super-Resolution with Events
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 EvEnhancer, an event-driven continuous space-time video super-resolution method combining event-adapted synthesis (EASM) and a local implicit video transformer (LIVT) for arbitrary spatial and temporal scales.
Problem
Continuous space-time video super-resolution (C-STVSR) struggles to produce quality video at out-of-distribution spatial and temporal scales, and existing methods do not exploit event camera streams for long-term motion modeling.
Approach
Uses event-modulated alignment and bidirectional recurrent compensation to model long-term motion trajectories from frames and events, then decodes a unified 3D video implicit neural representation via local cross-scale spatiotemporal attention instead of decoupled spatial and temporal INRs.
Key results
- In-distribution (t=4, s=4) on GoPro-Average: EvEnhancer reaches 33.30 dB / 0.9279 vs. HR-INR 32.13 dB / 0.9371 (higher PSNR, lower SSIM) and MoTIF 30.04 dB.
- On the OOD scale t=12, s=4 on GoPro, EvEnhancer reports 31.17 dB vs. MoTIF 27.77 dB and VideoINR 27.32 dB (the table lists EvEnhancer-light as 31.17 and EvEnhancer as 32.07).
- On 2x VSR on ALPIX-VSR, EvEnhancer achieves 40.84 dB / 0.9786 vs. EGVSR 39.31 dB / 0.9635.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the strongest event-based baseline HR-INR with EvEnhancer | C-STVSR · GoPro-Center, t=4, s=8 | PSNR (dB): 31.97 -> 33.52 | Helped |
| Replace the strongest event-based baseline HR-INR with EvEnhancer | C-STVSR · GoPro-Average, t=4, s=8 | PSNR (dB): 32.13 -> 33.30 | Helped |
| Replace the strongest event-based baseline HR-INR with EvEnhancer | C-STVSR · Adobe240-Center, t=4, s=8 | PSNR (dB): 31.26 -> 32.43 | Helped |
| Replace the strongest event-based baseline HR-INR with EvEnhancer | C-STVSR · Adobe240-Average, t=4, s=8 | PSNR (dB): 31.11 -> 32.18 | Helped |
| Use EvEnhancer instead of the strongest listed baseline at each scale | C-STVSR · GoPro OOD spatiotemporal scales in Table 2 | PSNR (dB) | Helped |
| Use EvEnhancer instead of the strongest listed baseline | C-STVSR · BS-ERGB, t=4, 6, and 8, s=4 | PSNR | Helped |
| Use EvEnhancer instead of EGVSR | 2× VSR · ALPIX-VSR | PSNR: 39.31 -> 40.84 | Helped |
| Add BRC to EMA (ablation) | EASM · GoPro In-Dist. (t=8, s=4) | PSNR (dB): 32.50 -> 33.13 | Helped |
| Add BRC to EMA (ablation) | EASM · GoPro OOD (t=12, s=6) | PSNR (dB): 28.86 -> 29.42 | Helped |
| Use forward-only rather than bidirectional EMA with BRC and BF (ablation) | EASM · GoPro In-Dist. (t=8, s=4) | PSNR (dB): 33.30 -> 33.20 | Hurt |
| Use forward-only rather than bidirectional EMA with BRC and BF (ablation) | EASM · GoPro OOD (t=12, s=6) | PSNR (dB): 29.54 -> 29.47 | Hurt |
| Replace 2D decoupled INR with unified 3D INR (ablation) | LIVT · GoPro In-Dist. (t=8, s=4) | PSNR (dB): 33.26 -> 27.32 | Hurt |
Surprising findings
- HR-INR achieves higher SSIM than EvEnhancer on in-distribution GoPro-Average (0.9371 vs 0.9279) despite lower PSNR.
Limitations
- MoTIF is computationally cheaper than EvEnhancer at low temporal scales (t<=4).
- HR-INR results are taken from its original paper because its source code is unavailable, so comparison conditions may differ.
- Training uses only the Adobe240 dataset with simulated events from vid2e.
Experimental setup
Datasets: Adobe240 GoPro BS-ERGB ALPIX-VSR
Benchmarks: Adobe240 GoPro BS-ERGB ALPIX-VSR
Metrics: PSNR SSIM TFLOPs
Baselines: TimeLens EGVSR EvTexture CBMNet-L REFID Zooming Slow-Mo TMNet SAFA VideoINR MoTIF HR-INR
Builds on: VideoINR MoTIF LIIF PCD alignment REFID
Closest papers by meaning
- Learning Spatial-Temporal Implicit Neural Representations for Event-Guided Video Super-Resolution (CVPR 2023 · similarity 0.91)
- Time Without Time: Pseudo-Temporal Representation for Space-Time Super-Resolution (CVPR 2026 · similarity 0.89)
- EvSTVSR: Event Guided Space-Time Video Super-Resolution (AAAI 2025 · similarity 0.89)
- Learning Trajectory-Aware Transformer for Video Super-Resolution (CVPR 2022 · 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.