Research map / Video / Video segmentation and tracking
KVQ: Boosting Video Quality Assessment via Saliency-guided Local Perception
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 KVQ, a Video Swin-T based no-reference VQA model with Fusion-Window Attention for saliency and a Local Perception Constraint for region-wise texture quality, and releases a 50-image LPVQ dataset with region-wise annotations.
Problem
No-reference video quality assessment that predicts global perceptual quality while also perceiving region-wise local quality, which is hard to supervise because region-level annotations are costly.
Approach
Decouples a video's quality into a saliency map (predicted via cross-window attention over correlated windows, Fusion-Window Attention) and a local texture map, and constrains the texture map so each region's prediction is consistent whether the region is fed alone or within the full video (Local Perception Constraint), without local supervision.
Key results
- KVQ reaches SRCC 0.896 / PLCC 0.897 on LSVQ test and SRCC 0.814 on LSVQ 1080p, improving over Fast-VQA by 2.3% and 4.5% SRCC.
- In transfer learning, KVQ improves YouTube-UGC SRCC to 0.903 versus 0.855 for Fast-VQA, and KoNViD-1k SRCC to 0.909.
- On the LPVQ local-quality benchmark, KVQ with LPC reaches inter-sample SRCC 0.614 versus 0.389 for Fast-VQA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace standard intra-window attention with Fusion-Window Attention combining intra- and cross-window attention (ablation) | Swin-T video VQA model · LSVQtest, LSVQ1080p, KoNViD-1k, LIVE-VQC | SRCC: 0.894 -> 0.896; 0.807 -> 0.814; 0.886 -> 0.890; 0.816 -> 0.820 | Helped |
| Use cross-window attention only rather than FWA (ablation) | Swin-T video VQA model · LSVQtest, LSVQ1080p, KoNViD-1k, LIVE-VQC | SRCC: 0.896 -> 0.881; 0.814 -> 0.793; 0.890 -> 0.871; 0.820 -> 0.801 | Hurt |
| Add the Local Perception Constraint (ablation) | KVQ · LSVQtest, LSVQ1080p, KoNViD-1k, LIVE-VQC | SRCC: 0.894 -> 0.896; 0.811 -> 0.814; 0.887 -> 0.890; 0.804 -> 0.820 | Helped |
| Use saliency-guided KVQ with LPC instead of Fast-VQA for local quality prediction | LPVQ · LPVQ inter-sample and intra-sample | SRCC: 0.389 -> 0.614; 0.280 -> 0.612 | Helped |
| Use KVQ rather than Faster-VQA for transfer learning | Models pretrained on LSVQ · KoNViD-1k | SRCC: 0.895 -> 0.909 | Helped |
| Use KVQ rather than Fast-VQA for no-reference VQA | Models trained on LSVQtrain · LIVE-VQC | SRCC: 0.823 -> 0.820 | Hurt |
Surprising findings
- Using only cross-window attention (CWA) performs worst among attention variants, indicating neighboring-region context is necessary for semantics.
Limitations
- LPVQ contains only 50 images annotated at the spatial level, used as static videos rather than true video local annotations.
- Saliency evaluation is zero-shot on SALICON, not trained for saliency, so saliency quality is only approximated.
Experimental setup
Datasets: LSVQ KoNViD-1k LIVE-VQC YouTube-UGC SALICON Kinetics-400 LPVQ
Benchmarks: LSVQ test LSVQ 1080p KoNViD-1k LIVE-VQC YouTube-UGC LPVQ SALICON
Metrics: SRCC PLCC sAUC NSS KL
Baselines: BRISQUE TLVQM VIDEVAL VSFA PVQ Li et al. CLIPVQA Fast-VQA Faster-VQA StarVQA VQT DisCoVQA
Builds on: Video Swin Transformer Fast-VQA PLCC loss Window attention
Closest papers by meaning
- Q-Bench-Video: Benchmark the Video Quality Understanding of LMMs (CVPR 2025 · similarity 0.87)
- KVQ: Kwai Video Quality Assessment for Short-form Videos (CVPR 2024 · similarity 0.87)
- PTM-VQA: Efficient Video Quality Assessment Leveraging Diverse PreTrained Models from the Wild (CVPR 2024 · similarity 0.87)
- Integrating Global Context Contrast and Local Sensitivity for Blind Image Quality Assessment (ICML 2024 · similarity 0.86)
- AIGV-Assessor: Benchmarking and Evaluating the Perceptual Quality of Text-to-Video Generation with LMM (CVPR 2025 · similarity 0.86)
- Knowledge Guided Semi-supervised Learning for Quality Assessment of User Generated Videos (AAAI 2024 · 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.