hollow404/MDS-VQA_sdr_diversity_features

MDS-VQA SDR Diversity Features

This dataset contains pre-extracted diversity features for the YouTube-SFV SDR videos used in MDS-VQA: Model-Informed Data Selection for Video Quality Assessment. The features are intended for the diversity term in the MDS-VQA greedy selection procedure.

Paper: arXiv:2603.11525
Project/code: Multimedia-Analytics-Laboratory/MDS-VQA

Dataset Summary

MDS-VQA selects unlabeled videos that are both difficult for a base video quality assessment model and diverse in content. Difficulty is predicted by a failure predictor, while diversity is computed from frame-level semantic video features. This repository provides the CLIP-based diversity features for YouTube-SFV SDR so users can run the MDS-VQA selection step without re-extracting these features.

The dataset includes feature files only. It does not include raw videos, MOS labels, failure scores, or base-model quality predictions.

Dataset Details

• Number of feature files: 2,030 .npz files
• Total repository size: approximately 51.6 MB
• Source video set: YouTube-SFV SDR
• Feature type: frame-level CLIP visual embeddings for diversity measurement
• Feature format: NumPy .npz
• Main use: --diversity-feature-path for MDS-VQA greedy selection
• License: Apache 2.0

The files follow the naming pattern:

SDR_<Category>_<VideoID>_features.npz

Example:

SDR_Animal_01ld_features.npz

Data Format

Each .npz file contains one array:

features.npy

The array stores frame-level visual features:

shape: (num_frames, 512)
dtype: float16

Usage with MDS-VQA

Download the feature repository:

from huggingface_hub import snapshot_download

feature_root = snapshot_download(
    repo_id="hollow404/MDS-VQA_sdr_diversity_features",
    repo_type="dataset",
)
print(feature_root)

Then pass the downloaded directory to the MDS-VQA selection script:

python src/greedy_selection.py \
  --video-dir /path/to/your/DATA_ROOT \
  --failure-score-path /path/to/your/failure_score/path \
  --baseline-pred-path /path/to/your/base-model_pred_quality_score/path \
  --diversity-feature-path /path/to/MDS-VQA_sdr_diversity_features \
  --model-name CLIP_RN101

The feature filenames should match the corresponding video identifiers expected by the selection script.

MDS-VQA Context

In MDS-VQA, the selected subset is optimized to balance:

1. Difficulty: high predicted failure score from the failure predictor g(.).
2. Diversity: high semantic dissimilarity between videos, measured from frame-level visual features.

The paper computes diversity using Chamfer distance between frame-level feature sets. The final greedy rule selects videos that are both likely to expose base-model errors and non-redundant with videos already selected.

Intended Uses

This dataset is intended for:

• reproducing the MDS-VQA greedy selection step on YouTube-SFV SDR;
• measuring diversity between candidate videos during active VQA data selection;
• avoiding repeated CLIP feature extraction for the provided SDR video pool;
• research on active learning, data selection, and video quality assessment.

Out-of-Scope Uses

This dataset is not a standalone VQA benchmark. It does not contain raw videos or human quality labels, and the feature embeddings are not sufficient to evaluate a VQA model by themselves. Users should obtain the corresponding raw videos and labels from the original data sources according to their terms.

Limitations

• The features are semantic visual embeddings, not direct quality features. They are designed to improve coverage during selection, not to predict perceptual quality alone.
• The features may not capture all perceptually important differences, especially motion-specific, temporal, audio, or metadata-driven failure modes.
• The feature files are tied to the YouTube-SFV SDR naming convention and should be aligned carefully with local video paths.
• Users should respect the license and terms of the original YouTube-SFV data when using raw videos.

Citation

If you use these features, please cite MDS-VQA:

@article{zou2026mds,
  title={MDS-VQA: Model-Informed Data Selection for Video Quality Assessment},
  author={Zou, Jian and Xu, Xiaoyu and Wang, Zhihua and Wang, Yilin and Adsumilli, Balu and Ma, Kede},
  journal={arXiv preprint arXiv:2603.11525},
  year={2026}
}