Research map / Video / Video segmentation and tracking
SAVSR: Arbitrary-Scale Video Super-Resolution via a Learned Scale-Adaptive Network
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 SAVSR, a single-model arbitrary-scale (non-integer and asymmetric) video super-resolution network using an iterative bi-directional alignment architecture, scale-attention convolution (OSConv), and spatio-temporal adaptive arbitrary-scale upsampling (STAU).
Problem
Existing video super-resolution networks only support fixed integer scale factors, requiring separate models per scale and failing at non-integer and asymmetric scales.
Approach
Combines an iterative window with bi-directional recurrent propagation for implicit alignment, makes convolution kernels adapt to the input scale via omni-dimensional scale attention, and upsamples with kernels generated from both spatio-temporal features and scale information.
Key results
- On UDM10 at x1.5 (asymmetric x1.5/x3.5 setting), SAVSR improves PSNR by 2.28 dB over the best arbitrary-scale image SR method ArbSR and by 0.36 dB over BasicVSR.
- On Vid4 at x2 symmetric scale, SAVSR reaches 32.43 dB/0.9351 vs BasicVSR 32.15 dB, showing competitive integer-scale performance.
- STAU achieves 27.65 dB at x3.7 with 121K extra params and 23G FLOPs, outperforming Meta-Upscale, LIIF-Upscale and LTE-Upscale upsamplers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the spatio-temporal filter branch to STAU (ablation) | SAVSR · Vid4, symmetric ×3.5 and asymmetric ×1.9/×3.5 | PSNR: 27.84 -> 28.03; 30.33 -> 30.58 | Helped |
| Add OSConv to alignment and reconstruction rather than using neither (ablation) | SAVSR · Vid4, symmetric ×2.25 and asymmetric ×1.3/×3.9 | PSNR: 31.78 -> 32.30; 29.88 -> 30.17 | Helped |
| Include scale factors in the OSConv scale vector (ablation) | SAVSR · Vid4, symmetric ×3.5 and asymmetric ×3.5/×1.5 | PSNR: 30.18 -> 30.23; 28.15 -> 28.22 | Helped |
| Use bidirectional rather than unidirectional propagation (ablation) | SAVSR iterative alignment architecture · Vid4, ×2.25 and ×2/×3.2 | PSNR: 31.84 -> 32.27; 30.75 -> 31.19 | Helped |
| Increase iteration window from 7 to 13 and sliding window from 3 to 5 (ablation) | SAVSR iterative alignment architecture · Vid4, ×2.25 and ×2/×3.2 | PSNR: 32.27 -> 32.33; 31.19 -> 31.25 | Helped |
Surprising findings
- Expanding the iteration and sliding window sizes does not proportionally improve PSNR but increases parameters.
Limitations
- Gains over BasicVSR at integer scales are small or absent (e.g., BasicVSR is better at some integer scales on Vid4 such as x4 by a small margin).
- Non-integer scale results rely on bicubic downsampling and bicubic post-processing for some baselines, which may not reflect real degradations.
- VideoINR performs poorly on the evaluated scales.
Experimental setup
Datasets: Vimeo-90K
Benchmarks: Vid4 UDM10
Metrics: PSNR SSIM
Baselines: Meta-SR LIIF ArbSR LTE DUF EDVR OVSR BasicVSR VideoINR
Builds on: Meta-SR LIIF BasicVSR EDVR CondConv
Closest papers by meaning
- Deep Video Super-Resolution Network Using Dynamic Upsampling Filters Without Explicit Motion Compensation (CVPR 2018 · similarity 0.90)
- Arbitrary-Scale Video Super-resolution Guided by Dynamic Context (AAAI 2024 · similarity 0.90)
- Accelerating the Training of Video Super-resolution Models (AAAI 2023 · similarity 0.90)
- Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network (CVPR 2016 · similarity 0.89)
- Frame-Recurrent Video Super-Resolution (CVPR 2018 · similarity 0.89)
- Sequence Matters: Harnessing Video Models in 3D Super-Resolution (AAAI 2025 · similarity 0.88)
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.