Research map / Video / Video-language understanding
KVQ: Kwai Video Quality Assessment for Short-form Videos
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.
Builds KVQ, a 4200-video short-form UGC quality database with MOS and pairwise rankings, and proposes KSVQE, a CLIP- and CONTRIQUE-guided 3D Swin Transformer VQA model that sets SOTA on KVQ and several UGC benchmarks.
Problem
Existing UGC video quality assessment methods fail on short-form UGC videos, where special effects and kaleidoscope content obscure quality-determined regions and sophisticated processing workflows create hard-to-distinguish hybrid distortions.
Approach
Use a CLIP-based quality adapter to select quality-aware patches (QRS) and modulate features with CLIP semantics (CaM), plus a CONTRIQUE-based distortion adapter trained with contrastive loss to modulate features with distortion priors (DaM), on top of a 3D Swin Transformer fragment backbone.
Key results
- On KVQ, KSVQE reaches SROCC 0.867 and PLCC 0.869, beating Dover* (0.833/0.837) by 0.034/0.032.
- On YouTube-UGC, KSVQE reaches SROCC 0.900 and PLCC 0.912, improving over the second-best by 0.059/0.061 as reported.
- Ablation on KVQ shows QRS, CaM and DaM each improve over the baseline (0.832/0.834), with QRS giving 0.015/0.019 gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KSVQE instead of the strongest reported baseline | VQA on KVQ · KVQ | SROCC: 0.833 -> 0.867 | Helped |
| Use KSVQE instead of the strongest reported baseline | VQA on KVQ · KVQ | PLCC: 0.837 -> 0.869 | Helped |
| Add quality-aware region selection (QRS) (ablation) | KSVQE · KVQ | SROCC: 0.832 -> 0.847 | Helped |
| Add content-adaptive modulation (CaM) (ablation) | KSVQE · KVQ | SROCC: 0.832 -> 0.842 | Helped |
| Add distortion-aware modulation (DaM) (ablation) | KSVQE · KVQ | PLCC: 0.834 -> 0.843 | Helped |
| Add both CaM and DaM instead of DaM alone (ablation) | KSVQE · KVQ | SROCC: 0.839 -> 0.850 | Helped |
| Add both CaM and DaM instead of CaM alone (ablation) | KSVQE · KVQ | SROCC: 0.842 -> 0.850 | Helped |
| Replace QRS with random region selection (ablation) | KSVQE · KVQ | SROCC: 0.847 -> 0.843 | Hurt |
| Replace quality-aware region selection with region selection (RS) (ablation) | KSVQE · KVQ | SROCC: 0.847 -> 0.843 | Hurt |
| Use cross-attention plus spatial modulation for CaM (ablation) | KSVQE · KVQ | SROCC: 0.835 -> 0.842 | Helped |
| Use cross-attention and self-attention plus channel modulation for DaM (ablation) | KSVQE · KVQ | SROCC: 0.830 -> 0.839 | Helped |
| Replace CaM with cross-attention only (ablation) | KSVQE · KVQ | SROCC: 0.842 -> 0.835 | Hurt |
Limitations
- Dataset is built from a single short-form platform (Kwai), which may limit generality.
- Distortion prior relies on CONTRIQUE pretrained on other distortion distributions and needs adaptation.
Experimental setup
Datasets: KVQ KoNViD-1k YouTube-UGC LSVQ LIVE-VQC
Benchmarks: KVQ KoNViD-1k YouTube-UGC LIVE-VQC LSVQ test LSVQ 1080p
Metrics: SROCC PLCC ranking accuracy
Baselines: TLVQM VIDEVAL VSFA GSTVQA SimpleVQA FastVQA DOVER PVQ RAPIQUE VIQE
Builds on: FastVQA DOVER CONTRIQUE CLIP 3D Swin Transformer
Closest papers by meaning
- KVQ: Boosting Video Quality Assessment via Saliency-guided Local Perception (CVPR 2025 · similarity 0.87)
- FineVQ: Fine-Grained User Generated Content Video Quality Assessment (CVPR 2025 · similarity 0.86)
- PTM-VQA: Efficient Video Quality Assessment Leveraging Diverse PreTrained Models from the Wild (CVPR 2024 · similarity 0.84)
- AIGV-Assessor: Benchmarking and Evaluating the Perceptual Quality of Text-to-Video Generation with LMM (CVPR 2025 · similarity 0.83)
- Koala-36M: A Large-scale Video Dataset Improving Consistency between Fine-grained Conditions and Video Content (CVPR 2025 · similarity 0.83)
- Q-Bench-Video: Benchmark the Video Quality Understanding of LMMs (CVPR 2025 · 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.