Research map / Video / Video-language understanding
VADB: A Large-Scale Video Aesthetic Database with Professional and Multi-Dimensional Annotations
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 VADB, a 10,490-video aesthetic dataset with professional multi-dimensional scores, comments and tags, and VADB-Net, a two-stage CLIP-based video-text pretraining plus MLP regression model for aesthetic scoring.
Problem
Video aesthetic assessment is limited by a lack of large, standardized, richly annotated datasets and by the difficulty of adapting image-based aesthetic methods to temporal, multimodal video data.
Approach
Pretrain a CLIP ViT-B/32-based video encoder contrastively against fused comment and aesthetic-tag text embeddings from a professionally annotated dataset, then freeze it and train a small MLP to regress overall and attribute aesthetic scores.
Key results
- VADB-Net reaches SRCC 0.93 on overall score on VADB, versus 0.92 for FastVQA, 0.89 for ModularBVQA and 0.85 for SimpleVQA.
- Removing the video-text pretraining stage (Ablation 1) lowers overall SRCC from 0.93 to 0.84.
- Overall-score statistical validation reports SROCC 0.9299 and binary accuracy 0.9180 with p<0.001.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use VADB-Net instead of the strongest baseline, FastVQA | VADB-Net · VADB overall | SRCC: 0.92 -> 0.93 | Helped |
| use VADB-Net instead of the strongest baseline, FastVQA | VADB-Net · VADB overall | PLCC: 0.93 -> 0.93 | No clear effect |
| use VADB-Net instead of the strongest baseline, FastVQA | VADB-Net · VADB overall | RMSE: 0.65 -> 0.54 | Helped |
| use VADB-Net instead of the strongest baseline, FastVQA | VADB-Net · VADB overall | KRCC: 0.76 -> 0.77 | Helped |
| add VADB pretraining (ablation) | VADB-Net · VADB overall | SRCC: 0.84 -> 0.93 | Helped |
| add VADB pretraining (ablation) | VADB-Net · VADB overall | PLCC: 0.85 -> 0.93 | Helped |
| add VADB pretraining (ablation) | VADB-Net · VADB overall | RMSE: 0.59 -> 0.54 | Helped |
| add VADB pretraining (ablation) | VADB-Net · VADB overall | KRCC: 0.65 -> 0.77 | Helped |
| use the two-hidden-layer MLP instead of a linear layer (ablation) | VADB-Net · VADB overall | SRCC: 0.90 -> 0.93 | Helped |
| use the two-hidden-layer MLP instead of a linear layer (ablation) | VADB-Net · VADB overall | PLCC: 0.90 -> 0.93 | Helped |
| use the two-hidden-layer MLP instead of a linear layer (ablation) | VADB-Net · VADB overall | RMSE: 0.51 -> 0.54 | Hurt |
| use the two-hidden-layer MLP instead of a linear layer (ablation) | VADB-Net · VADB overall | KRCC: 0.72 -> 0.77 | Helped |
Surprising findings
- Replacing the MLP head with a linear layer (Ablation 2) gives a slightly lower RMSE on the overall score (0.51) than the full VADB-Net (0.54), despite lower SRCC.
Limitations
- Annotators are all from one institution (Beijing Film Academy), introducing single-cultural bias.
- Dataset focuses on human-centered videos (character category dominates), possibly causing human-centric aesthetic bias.
- Inter-annotator agreement is moderate (Krippendorff's alpha about 0.54 to 0.66), reflecting subjectivity.
Experimental setup
Datasets: VADB
Benchmarks: VADB
Metrics: SRCC PLCC RMSE KRCC MSE binary accuracy
Baselines: FAST-VQA SimpleVQA ModularBVQA
Builds on: CLIP CLIP4Clip FAST-VQA SimpleVQA ModularBVQA
Closest papers by meaning
- Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.87)
- ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding (CVPR 2026 · similarity 0.86)
- Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective. (ICLR 2026 · similarity 0.86)
- APDDv2: Aesthetics of Paintings and Drawings Dataset with Artist Labeled Scores and Comments (NeurIPS 2024 · similarity 0.86)
- Advancing Comprehensive Aesthetic Insight with Multi-Scale Text-Guided Self-Supervised Learning (AAAI 2025 · similarity 0.84)
- VisJudge-Bench: Aesthetics and Quality Assessment of Visualizations (ICLR 2026 · 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.