Research map / Video / Video segmentation and tracking
FineVQ: Fine-Grained User Generated Content Video Quality Assessment
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.
Introduces FineVD, a 6104-video UGC quality dataset with multi-dimensional MOS, degradation labels and descriptions, and FineVQ, an LMM-based model (InternViT + SlowFast motion encoder + InternLM-8B with LoRA) that rates, scores and describes quality per dimension.
Problem
Existing UGC video quality assessment databases and models provide only an overall quality score, lacking fine-grained, multi-dimensional quality labels, attributions and descriptions needed for processing and recommendation.
Approach
Builds a lab-annotated multi-dimensional UGC VQA dataset with MOS, distortion-type and description labels, and an instruction-tuned large multimodal model with spatial and motion encoders and LoRA adaptation that handles quality rating, scoring and attribution in a single model.
Key results
- On FineVD quality scoring, FineVQ reaches overall SRCC 0.8834 and PLCC 0.8891, best across all six dimensions versus DOVER and SimpleVQA.
- On LSVQ 1080p, FineVQ achieves SRCC 0.828 and PLCC 0.857, a reported 3.8% improvement over KSVQE.
- In quality attribute prediction, FineVQ outperforms seven open LMMs on FineVD, e.g., 91.93% accuracy on the 'Most' impact question versus 43.21% for the best baseline (InternVL2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add motion features to spatial features (ablation) | FineVQ · FineVD; LIVE-VQC; YouTube-UGC | SRCC: 0.8210 -> 0.8264; 0.7832 -> 0.8061; 0.8084 -> 0.8349 | Helped |
| add vision-encoder LoRA rank 8 (ablation) | FineVQ with spatial and motion features · FineVD; LIVE-VQC; YouTube-UGC | SRCC: 0.8264 -> 0.8345; 0.8061 -> 0.8207; 0.8349 -> 0.8442 | Helped |
| add language-model LoRA rank 8 (ablation) | FineVQ with spatial and motion features · FineVD; LIVE-VQC; YouTube-UGC | SRCC: 0.8264 -> 0.8365; 0.8061 -> 0.8224; 0.8349 -> 0.8640 | Helped |
| use LoRA rank 8 on both vision encoder and language model (ablation) | FineVQ · FineVD; LIVE-VQC; YouTube-UGC | SRCC: 0.8365 -> 0.8838; 0.8224 -> 0.8624; 0.8640 -> 0.8921 | Helped |
| increase LoRA rank from 8 to 16 (ablation) | FineVQ with both LoRA adaptations · FineVD; LIVE-VQC; YouTube-UGC | SRCC: 0.8838 -> 0.8887; 0.8624 -> 0.8951; 0.8921 -> 0.9107 | Helped |
| use one FineVQ model for quality attribution instead of the compared general LMMs | FineVQ · FineVD quality attribute prediction | accuracy: 91.93% | Helped |
Surprising findings
- Several general LMMs score below random chance on yes-or-no distortion detection (e.g., Video-LLaVA 26.48% on color vs 50% chance).
- Models trained on FineVD generalize better to other datasets than the reverse, suggesting FineVD is more challenging.
Limitations
- Weaker performance on the temporal dimension, which the authors attribute to insufficient temporal features.
- Dataset annotations come from 22 in-lab annotators, which limits scale relative to crowdsourced datasets.
- Runtime is higher than some lightweight VQA baselines such as FastVQA and DOVER.
Experimental setup
Datasets: FineVD KoNViD-1k YouTube-UGC LIVE-VQC LSVQ LIVE-YT-Gaming
Benchmarks: FineVD LIVE-YT-Gaming KoNViD-1k YouTube-UGC LIVE-VQC LSVQ test LSVQ 1080p
Metrics: SRCC PLCC KRCC accuracy
Baselines: NIQE QAC HOSA TLVQM VIDEVAL RAPIQUE VSFA GSTVQA SimpleVQA FAST-VQA DOVER VideoLLaMA2
Builds on: LoRA InternVL SlowFast instruction tuning
Closest papers by meaning
- KVQ: Kwai Video Quality Assessment for Short-form Videos (CVPR 2024 · similarity 0.86)
- PTM-VQA: Efficient Video Quality Assessment Leveraging Diverse PreTrained Models from the Wild (CVPR 2024 · similarity 0.85)
- Modular Blind Video Quality Assessment (CVPR 2024 · similarity 0.85)
- AIGV-Assessor: Benchmarking and Evaluating the Perceptual Quality of Text-to-Video Generation with LMM (CVPR 2025 · similarity 0.85)
- Knowledge Guided Semi-supervised Learning for Quality Assessment of User Generated Videos (AAAI 2024 · similarity 0.84)
- UltraVideo: High-Quality UHD Video Dataset with Comprehensive Captions (NeurIPS 2025 · similarity 0.84)
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.