README.md

metadata

license: unknown
pretty_name: Stanford I2V 600K
task_categories:
  - image-to-video
tags:
  - mteb
  - moeb
  - image-to-video-audio-retrieval
configs:
  - config_name: corpus
    data_files:
      - split: test
        path: corpus/test-*
  - config_name: qrels
    data_files:
      - split: test
        path: qrels/test-*
  - config_name: queries
    data_files:
      - split: test
        path: queries/test-*
dataset_info:
  - config_name: corpus
    features:
      - name: id
        dtype: string
      - name: video
        dtype: video
      - name: audio
        dtype: audio
    splits:
      - name: test
        num_bytes: 63291770108
        num_examples: 3401
    download_size: 71073630159
    dataset_size: 63291770108
  - config_name: qrels
    features:
      - name: query-id
        dtype: string
      - name: corpus-id
        dtype: string
      - name: score
        dtype: int64
    splits:
      - name: test
        num_bytes: 99186
        num_examples: 1280
    download_size: 21358
    dataset_size: 99186
  - config_name: queries
    features:
      - name: id
        dtype: string
      - name: image
        dtype: image
    splits:
      - name: test
        num_bytes: 14665003
        num_examples: 229
    download_size: 14669276
    dataset_size: 14665003

Stanford I2V 600K

Frozen MTEB/MOEB representation of the official Stanford I2V 600K evaluation release for image-to-video+audio retrieval.

Construction

The dataset is reconstructed from Stanford Digital Repository record zx935qw7203, version 15. Query IDs, corpus paths, and relevance annotations are preserved from 600k_dataset_queries_public.txt, list_database_clips_600k_public.txt, and 600k_dataset_public.txt. Multiple temporal annotations for the same query/video pair are collapsed to one binary scene-level relevance judgment, matching the source scene-retrieval protocol.

The construction script is maintained in the MTEB repository at scripts/data/stanford_i2v_retrieval/create_data.py. Every source archive part is checked against the size and MD5 digest in the Stanford record before extraction. Images are decoded, videos are probed for a valid video stream, and their AAC soundtracks are placed in a separate audio column and fully decoded with strict error handling. Soundtracks with valid AAC streams are losslessly remuxed. The 81 tracks containing malformed source packets are decoded and re-encoded while preserving timestamp gaps. One of those tracks has an unrecoverable first 69 seconds, which are replaced with silence; its only relevance annotation (2:25-2:27) lies in the recoverable suffix.

Frozen evaluation contents

• Queries: 229 (official: 229)
• Corpus video+audio clips: 3401 (official: 3401)
• Binary query/video qrels: 1280 (official: 1280)
• Source temporal annotations: 2260
• Distinct positive corpus videos: 1188 (official: 1188)
• Distractor corpus videos: 2213 (official: 2213)
• Missing or corrupt queries: 0
• Missing or corrupt videos: 0
• Missing or corrupt audio tracks: 0
• Re-encoded source-malformed audio tracks: 81
• Tracks with an unrecoverable prefix replaced by silence: 1
• Queries removed because no positive survived: 0

Configs follow MTEB's standard retrieval representation: queries, corpus, and qrels, each with a test split.

Evaluation

The source benchmark reports mean average precision over the top 100 retrieved scenes. The corresponding MTEB primary metric is map_at_100. Stanford's scorer divides each query's precision sum by the total number of relevant corpus videos, which matches standard truncated AP@100 for this release.

Comparability with the 2015 paper

The 600K query and ground-truth manifests are byte-identical to the full release: all 229 queries and all annotated positive videos are preserved. The corpus is the official 3,401-video 600K corpus rather than the 84,443-video full corpus used for the principal 2015 paper results. The query set, qrels, and metric are therefore identical, but scores are not directly comparable because the 600K release contains fewer distractor videos. The 600K release was documented in the later Bloom-filter retrieval paper.

Licensing and provenance

The Stanford source record provides public download access but does not specify a dataset-wide license. Accordingly, this repository records the license as not specified and does not claim that the source media is under an open license. Query images originate from news websites and corpus videos are recorded newscasts; rights may remain with their respective owners. Users are responsible for confirming that their use complies with applicable source terms and law.

Citation

@inproceedings{AraujoMMSYS2015,
  author = {Araujo, A. and Chaves, J. and Chen, D. and Angst, R. and Girod, B.},
  booktitle = {Proc. ACM Multimedia Systems},
  title = {Stanford I2V: A News Video Dataset for Query-by-Image Experiments},
  year = {2015},
}

@article{AraujoArxiv2016,
  author = {Araujo, A. and Chaves, J. and Lakshman, H. and Angst, R. and Girod, B.},
  journal = {arXiv preprint arXiv:1604.07939},
  title = {Large-Scale Query-by-Image Video Retrieval Using Bloom Filters},
  year = {2016},
}