EVVE Video-to-Video Event Retrieval

This repository packages a fixed, self-contained evaluation set derived from the EVent VidEo (EVVE) benchmark introduced by Revaud et al. at CVPR 2013. Given a query video of a specific real-world event, the task is to retrieve other videos depicting the same event.

Original and packaged protocols

The published EVVE core contains 620 queries, 2,375 database videos, 13 events, 135,213 positive query/database judgments, and 166 hours of web video. The original large-scale experiment additionally used 100,000 distractor videos.

The public S2VS EVVE feature artifact identifies a surviving 2,410-video layer of the original core. A reproducible acquisition pass obtained 2,110 of those videos from public media sources: 2,042 directly from their original YouTube IDs and 68 from checksum-pinned public archive captures (67 Wayback Machine captures and one Internet Archive item). Intersecting the 2,110-video manifest with the checksum-pinned annotation snapshot yields:

• 466 query videos (154 fewer than the original core);
• 1,644 database videos (731 fewer than the original core);
• 13 events;
• 86,925 positive query/database relevance judgments;
• 958 database videos relevant to at least one event and 686 other core candidates;
• every retained query has at least one positive;
• zero query/corpus ID overlap.

The acquisition filter removed 300 IDs from the S2VS layer: 38 queries and 262 database videos (81 positive and 181 other database videos), eliminating 13,864 qrels. At freeze time, yt-dlp reported 95 private videos, 64 videos from terminated accounts, 14 age-restricted videos, four copyright-claimed or blocked videos, one terms-of-service removal, and 122 otherwise unavailable videos. Exact attrition is reproducible as the set difference between construction/s2vs-2023-video-ids.txt (2,410 IDs) and construction/surviving-public-media-video-ids.txt (2,110 IDs).

media-manifest.jsonl records the exact frozen ID, packaged filename, original and acquisition URLs, byte count, and SHA-256 checksum for every video. The manifest's own SHA-256 is 9ef60f6584b0ca6b2cfe9be6ba982d10b06cb37fc1d8ee52718a5f0349efa8fb. Construction fails if any frozen video is missing or contains no decodable frames, so later source availability cannot silently alter the task. The complete frozen tree also passed MTEB's TorchCodec 10-frame uniform sampler. media-audit.json records final completeness and affected-qrel counts.

Event coverage

Q/P means retained queries / positive database videos. All 13 official events remain represented, and every packaged query has one or more packaged positives.

Event	Original Q/P	S2VS Q/P	Packaged Q/P	Removed from S2VS Q/P
barcelona+riots+2012	13/27	11/17	9/13	2/4
die+toten+hosen+rock+am+ring+2012	32/64	21/42	19/37	2/5
dsk+arrested	9/19	8/17	8/17	0/0
egypt+Tahrir+Square+protestors	36/72	32/62	31/59	1/3
johnny+stade+de+france+2012	87/174	75/155	75/150	0/5
kate+william+wedding	44/88	30/61	25/55	5/6
marrakech+bomb+attack	4/10	3/8	3/6	0/2
mdna+rome+2012	51/104	38/86	31/76	7/10
obama+speech+victory	14/29	11/22	10/21	1/1
shakira+live+kiev+2011	19/39	10/24	8/23	2/1
strokkur+geyser	215/431	189/387	175/357	14/30
thailand+flood+2011	73/148	55/115	52/106	3/9
universal+studios+jurassic+park+ride	23/47	21/43	20/38	1/5

Construction and relevance

The source metadata identifies videos by YouTube ID. The construction script downloads video at up to 360p with yt-dlp, without transcoding. Four IDs whose progressive MP4 currently contains no media payload use separate complete video-only and audio streams, which yt-dlp merges without transcoding. For source IDs no longer downloadable from YouTube, checksum-pinned public archive captures in construction/media-source-overrides.json are used. The original and actual acquisition URLs remain separate in every row and in media-manifest.jsonl. The resulting media is embedded in the Hugging Face dataset. The corpus, queries, and qrels configurations use the standard MTEB retrieval schema. Query rows also retain the EVVE event name and event-specific ignored (null) database IDs required by the original evaluator (the pinned annotation snapshot contains no retained null IDs).

For every retained query, positive qrels are the Cartesian product with all retained positive database videos for the same official EVVE event. The 100,000-video distractor collection is not included because the original distractor media is no longer publicly available.

This is a reproducible surviving-public-media subset of EVVE, not the complete original benchmark. Because both the core corpus and distractor protocol differ, published/source-paper scores are not directly comparable with results here.

Media provenance and license

The media originated from the YouTube IDs published by EVVE; public archival captures are used only where listed explicitly in the source-override manifest. No applicable dataset-wide redistribution license for the original videos was identified; the archived project page also contains no explicit third-party redistribution prohibition. Copyright remains with the respective video rights holders. The license is therefore recorded as not specified. This provenance statement does not claim that video copyright was transferred to EVVE or to this repository, or that redistribution is authorized.

The EVVE project page states separately that its software is BSD-licensed and its released descriptors are free. Those terms are not presented here as a license for the packaged videos.

Sources

• Archived EVVE project page
• Original paper
• Pinned S2VS annotations
• S2VS EVVE feature artifact

Citation

@inproceedings{revaud2013event,
  author = {Revaud, Jerome and Douze, Matthijs and Schmid, Cordelia and Jegou, Herve},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages = {2459--2466},
  title = {Event Retrieval in Large Video Collections with Circulant Temporal Encoding},
  year = {2013}
}