pranitchawla/VCDB-Core

VCDB Core Retrieval

This repository packages the 528-video core set of VCDB as a symmetric video-to-video retrieval task for MTEB/MOEB. It does not include VCDB's separate 100,000-video background collection. For each query, every other core video not marked relevant acts as an implicit distractor.

Terms and provenance

The source dataset is provided by Fudan University for research purposes only. The source authors and Fudan University make no warranties about the dataset, including non-infringement. Users must review and follow the original VCDB terms. The original page is https://fvl.fudan.edu.cn/dataset/vcdb/list.htm and may be intermittently unavailable.

The videos were collected from YouTube and MetaCafe. Rights in the underlying videos remain with their respective rightsholders. This packaging does not grant additional rights beyond the source terms.

Construction

• Raw core videos: 528
• Raw annotation groups: 28
• Raw segment annotations: 9236
• Retained segment annotations: 9233
• Query videos: 528
• Corpus videos: 528
• Unique undirected relevant video pairs: 5610
• Directed binary qrels: 11220

Multiple copied segments between the same parent-video pair are collapsed into one binary judgment. Every pair is emitted in both directions. Query and corpus IDs match for the same physical video; the MTEB task removes this self-match before scoring. The 528 diagonal segment annotations are retained in annotations but do not create self-relevance judgments.

One source video is stored as .mp4 although six annotation references name the same group-local hash with a .flv suffix. These references are resolved to its actual topic-relative archive path; the audit records every occurrence.

All timestamps have valid HH:MM:SS syntax, positive segment lengths, and non-negative values. The audit records 38 retained endpoints that exceed the duration reported by their archived media container. They are preserved because they are present in the official annotations, are not among the three rows excluded by the public reference setup, and do not affect the video-level relationship.

The complete retained segment-level timestamps are preserved in the annotations configuration. Three rows removed by the public facebookresearch VCDB setup are excluded:

• mr_and_mrs_smith_tango.txt:21: 38f11a4162d8e94227ac644f117f942735b9a504.mp4,bf582249cfc79d691195a8681961029cc5149a76.flv,00:02:35,00:03:15,00:01:04,00:03:18
• obama_kicks_door.txt:234: 14c262ea09b4ca66feb7e88cf57e0faaeacc301f.mp4,f150e062960b477adcac3f12ef4543337f5a91a4.flv,00:00:19,00:00:33,00:00:00,00:00:14
• run_forrest_fun.txt:5: 067fb2aa9623905a42a2b0b286de1386e45c5bf8.flv,8f19329946455ae5c2c7b788ea6f6513bf5e1c9a.flv,00:01:06,00:02:18,00:00:05,00:01:17

Media conversions

None; all source videos were preserved byte-for-byte.

Any conversion uses single-threaded FFmpeg/libx264, CRF 18, the medium preset, yuv420p, no audio, no metadata or chapters, and faststart.

Configurations

• corpus (test): id: string, video: Video
• queries (test): id: string, video: Video
• qrels (test): query-id: string, corpus-id: string, score: int32
• annotations (test): source group/row, both video IDs, and both time ranges

Source archive SHA-256

• core_dataset.z01: cc4b2f54ee6e13277c7657061c28b101b97f95e5e3897f1246bab0c432d70ddf (730000000 bytes)
• core_dataset.z02: a1b99131f374c841bb79906704e64555119ffa2307bc6915ac5b843dcabdddc3 (730000000 bytes)
• core_dataset.z03: c34b0119bd136d02c9d288e53a8546cfe2f477f52f459fc7a483313ec32ec394 (730000000 bytes)
• core_dataset.z04: 05c591e74e0e51a05703c6c7b93d6e46bf0c4e80e8ab94b0f1a5ceae688046a5 (730000000 bytes)
• core_dataset.z05: fe346c17fb971833b5f4dbef8c3313d84e47d8d0f6c00f5165b7be6ca65bbf3e (730000000 bytes)
• core_dataset.z06: 01973a3b5eab4ef8451ac6f2434a71d14aff8ebd2c1b0af078dbfd20672cf522 (730000000 bytes)
• core_dataset.z07: 8f700e059bf29b1dfdde4cb9bc6c38b1ad1ba12152c70617a2116e732eb66cb6 (730000000 bytes)
• core_dataset.z08: 6bb5286b05987de11a076b4fd0985017f4ca931bc53cc1cd16510d802690f0a3 (730000000 bytes)
• core_dataset.z09: 84ba7436187dc4bc41ce21475b3f70b5887dd800da8e61beea660650dd322d76 (730000000 bytes)
• core_dataset.z10: 31e9710787b5994118304a9f336312943c81b643a56e5e2425577755d7362f7c (730000000 bytes)
• core_dataset.z11: 7164ad7d655398e092580fdcda734dbead0960e8dee858a62e8dd4218bbc05c4 (730000000 bytes)
• core_dataset.zip: 22f2e052e835f30e4bd1cc794ff2c1a564923fed7d5fa96d8db5afca715a5a16 (260421905 bytes)

Limitations

VCDB was designed for temporal partial-copy localization. This adaptation evaluates whole-video retrieval and therefore cannot reproduce the paper's segment-level F1 score. It also uses only the core set, giving a 528-video corpus rather than the complete 100K-background protocol.

Citation

@inproceedings{jiang2014vcdb,
  author = {Jiang, Yu-Gang and Jiang, Yudong and Wang, Jiajun},
  booktitle = {Computer Vision -- ECCV 2014},
  pages = {357--371},
  publisher = {Springer},
  title = {VCDB: A Large-Scale Database for Partial Copy Detection in Videos},
  year = {2014},
}

Paper: https://fvl.fudan.edu.cn/_upload/article/files/7b/c4/190424104d2192e8e83cb9dfa6fc/5a8d85f7-5738-450c-b957-3de71d8d7e72.pdf