Research map / Video / Video segmentation and tracking
Knowledge Guided Semi-supervised Learning for Quality Assessment of User Generated 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.
Proposes ST-VQRL, a self-supervised Video Swin-T pretraining with a statistical (MVG-distance) contrastive loss, and SSL-VQA, a dual-model semi-supervised framework that transfers knowledge between a regressor and a pristine-distance model for no-reference UGC video quality assessment.
Problem
No-reference UGC video quality assessment requires large human-annotated datasets, which are costly; the paper targets VQA with very limited labels plus unlabelled videos.
Approach
Self-supervised pretraining with a statistical contrastive loss based on multivariate Gaussian distances between clip features, followed by semi-supervised training of two quality models that exchange pseudo-labels, gated by which model is more stable under augmentation.
Key results
- SSL-VQA^- (labelled only, 500 labels) reaches SROCC 0.719 on LSVQ test and 0.736 on KoNViD-1K, outperforming FAST-VQA (0.682 and 0.679).
- Full SSL-VQA with 1500 unlabelled videos reaches SROCC 0.765 on KoNViD-1K and 0.711 on LIVE VQC, above FixMatch (0.727 and 0.685).
- Ablation: removing the knowledge transfer loss drops KoNViD-1K SROCC from 0.765 to 0.742.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SSL-VQA instead of FixMatch with the ST-VQRL backbone | ST-VQRL-based VQA · KoNViD-1K (labeled and unlabeled data) | SROCC: 0.727 -> 0.765 | Helped |
| Add the statistical contrastive loss instead of similarity loss (ablation) | ST-VQRL · KoNViD-1K | SROCC: 0.733 -> 0.765 | Helped |
| Add the statistical contrastive loss instead of similarity loss (ablation) | ST-VQRL · LIVE VQC | SROCC: 0.684 -> 0.711 | Helped |
| Replace pretrained Video Swin-T with ST-VQRL using statistical loss (ablation) | SSL-VQA backbone · KoNViD-1K | SROCC: 0.728 -> 0.765 | Helped |
| Replace pretrained Video Swin-T with ST-VQRL using statistical loss (ablation) | SSL-VQA backbone · LIVE VQC | SROCC: 0.686 -> 0.711 | Helped |
| Add intra-model consistency loss (ablation) | SSL-VQA · KoNViD-1K | SROCC: 0.756 -> 0.765 | Helped |
| Add intra-model consistency loss (ablation) | SSL-VQA · LIVE VQC | SROCC: 0.697 -> 0.711 | Helped |
| Add knowledge transfer loss (ablation) | SSL-VQA · KoNViD-1K | SROCC: 0.742 -> 0.765 | Helped |
| Add knowledge transfer loss (ablation) | SSL-VQA · LIVE VQC | SROCC: 0.686 -> 0.711 | Helped |
| Add consistency loss and its stability mask (ablation) | SSL-VQA− · LSVQ_test | SROCC: 0.704 -> 0.719 | Helped |
| Increase the number of labeled training videos (ablation) | SSL-VQA · KoNViD-1K, LIVE VQC, LSVQ_test, and LSVQ_1080p | SROCC | Helped |
| Increase unlabeled training videos beyond 1500 (ablation) | SSL-VQA · LSVQ-based evaluation | SROCC | No clear effect |
Surprising findings
- Pre-trained Video Swin-T with supervised action-recognition pretraining performs worse than ST-VQRL trained from scratch on quality-specific self-supervised objectives (0.728 vs 0.765 on KoNViD-1K).
Limitations
- Semi-supervised experiments use a small labelled subset (500 videos) and a single unlabelled pool from LSVQ, with results medians over only 3 splits.
- Pristine reference corpus of 60 videos from multiple datasets is required for the distance-based model.
- Mean Teacher, FixMatch and Meta Pseudo-Label were adapted by the authors to regression, so baseline tuning may differ from originals.
Experimental setup
Datasets: LSVQ LIVE-VQA LIVE Mobile CSIQ VQD EPFL-PoLiMI ECCV-EVVQ KoNViD-1K LIVE VQC YouTube-UGC LIVE Qualcomm
Benchmarks: LSVQ test LSVQ 1080p KoNViD-1K LIVE VQC YouTube-UGC LIVE Qualcomm
Metrics: SROCC PLCC
Baselines: VBLIINDS TLVQM VIDEVAL VSFA FAST-VQA TCSVT-BVQA VISION CONVIQT Mean Teacher FixMatch Meta Pseudo-Label
Builds on: Video Swin Transformer VISION FAST-VQA NIQE FixMatch
Closest papers by meaning
- PTM-VQA: Efficient Video Quality Assessment Leveraging Diverse PreTrained Models from the Wild (CVPR 2024 · similarity 0.88)
- KVQ: Boosting Video Quality Assessment via Saliency-guided Local Perception (CVPR 2025 · similarity 0.86)
- Semi-Supervised Blind Image Quality Assessment through Knowledge Distillation and Incremental Learning (AAAI 2024 · similarity 0.85)
- AIGV-Assessor: Benchmarking and Evaluating the Perceptual Quality of Text-to-Video Generation with LMM (CVPR 2025 · similarity 0.85)
- FineVQ: Fine-Grained User Generated Content Video Quality Assessment (CVPR 2025 · similarity 0.84)
- Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking (ICLR 2026 · 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.