CoChat

VCDB: A Large-Scale Database for Partial Copy Detection in Videos

Jiang, Yu-Gang, Jiang, Yudong, Wang, Jiajun · 2014 · Lecture Notes in Computer Science

The DOI matches its Crossref record; 24 of the 24 references it cites were found in a registry or an index; 3 of the 14 sources its claims cite were checked for support.

Overview

CoChat Summary

TL;DR

VCDB is introduced as a large-scale video copy database with over 100,000 Web videos and 9,236 manually annotated real partial copy pairs, addressing the lack of realistic partial-copy benchmarks. Benchmarking a state-of-the-art baseline shows that methods with near-perfect results on simulated datasets perform poorly on complex real copies.

Key findings

  • VCDB contains 528 core videos (approximately 27 hours) from 28 queries, plus 100,000 YouTube background distraction videos; annotation found 9,236 pairs of partial copies.
  • Among annotated copies, 36% contain insertion of patterns, 18% are from camcording, 27% have scale changes, and 2% contain picture-in-picture; nearly 80% of video pairs with at least one copy contain just one copied segment, while 20% contain two or more.
  • 32% of copied segments are less than 10 seconds and another 28% are between 10 and 30 seconds; 44% of copies are shorter than 1/5 of their parent videos, and only 31% occupy over 80% of the parent videos.
  • On Muscle-VCD ST2, the baseline achieves QS=0.81 and QF=0.70, versus best reported 0.86 and 0.76. On VCDB core, temporal network alignment reaches segment-level recall around 0.48 at precision 0.80; frame-level recall saturates at 0.5 for temporal network and 0.45 for Hough voting, and best recall is close to 0.60 when precision drops to 0.20.

Methods

  • Dataset construction: videos downloaded from YouTube and MetaCafe, starting from 28 queries with about 20 manually picked videos per query, yielding 528 core videos; 100,000 YouTube distraction videos were added. Seven trained annotators used a custom tool with transitivity propagation over about one month (around 700 man-hours).
  • Baseline pipeline: uniform sampling of two frames per second; local SIFT descriptors; bag-of-visual-words quantization via hierarchical k-means; inverted file indexing with Hamming embedding; weak geometric consistency; temporal alignment compared via temporal network optimization and temporal Hough voting.
  • Evaluation: Muscle-VCD ST2 with 21 query segments using QF and QS; VCDB with each of the 9,236 pairs as a query using segment-level and frame-level precision/recall; large-scale tests incrementally adding 10,000, 30,000, 50,000, and 100,000 distraction videos.

Limitations

  • The paper states that manual annotation of real partial copies is extremely difficult and time-consuming, and that the transitivity property used to speed annotation does not always hold, e.g., when a bridging segment contains picture-in-picture scenes.
  • The paper notes the baseline uses standard weak geometric consistency, whereas the best Muscle-VCD result used an improved version; it also samples only two frames per second and says evaluating more frames is beyond the paper’s focus.
  • Evident from the text: the core dataset is built from 28 queries and YouTube/MetaCafe videos, and distraction videos were only skimmed to reduce the chance of copies, so coverage of all real partial copies is not guaranteed; the benchmark uses one representative baseline rather than an exhaustive comparison.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/978-3-319-10593-2_24

Cites

24 of 24 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Three pairs of frames extracted from copied video segments in VCDB. All the copies were found directly from the Internet through careful manual annotation. The complex forms of transformations in VCDB pose new challenges to video copy detection research, as the existing datasets were mostly generated 'artificially' by imposing a very few number of pre-defined transformations.
Fig. 1 Three pairs of frames extracted from copied video segments in VCDB. All the copies were found directly from the Internet through careful manual annotation. The complex forms of transformations in VCDB pose new challenges to video copy detection research, as the existing datasets were mostly generated 'artificially' by imposing a very few number of pre-defined transformations.From the PDF
Fig. 2 An example of a video pair containing multiple partial copies. Similar cases are frequently seen in VCDB.
Fig. 2 An example of a video pair containing multiple partial copies. Similar cases are frequently seen in VCDB.From the PDF
Fig. 3 Example frame copies from the videos downloaded by the 28 queries, respectively. Ordered from left to right and top to bottom, the corresponding queries are topics about commercials (3), movies (11), music video (1), public speeches (3), sports (6), surveillance event (1), and others (3).
Fig. 3 Example frame copies from the videos downloaded by the 28 queries, respectively. Ordered from left to right and top to bottom, the corresponding queries are topics about commercials (3), movies (11), music video (1), public speeches (3), sports (6), surveillance event (1), and others (3).From the PDF
Fig. 4 Statistics of VCDB: (a) the number of partial copies per video pair, among those having at least one copy; (b) the duration of the partial copies; and (c) the percentage of the duration of the copy segments in the corresponding parent videos. See texts for more explanations.
Fig. 4 Statistics of VCDB: (a) the number of partial copies per video pair, among those having at least one copy; (b) the duration of the partial copies; and (c) the percentage of the duration of the copy segments in the corresponding parent videos. See texts for more explanations.From the PDF
Fig. 5 The general framework of a video copy detection system.
Fig. 5 The general framework of a video copy detection system.From the PDF
Fig. 6 Precision-recall curves of the baseline system on the core dataset of VCDB, using the two temporal alignment methods respectively. Left: segment-level results. Right: frame-level results. Overall, the performance is much worse than that reported on the existing datasets with simulated copies, indicating that partial copy detection in realistic videos remains a challenging problem that deserves future research.
Fig. 6 Precision-recall curves of the baseline system on the core dataset of VCDB, using the two temporal alignment methods respectively. Left: segment-level results. Right: frame-level results. Overall, the performance is much worse than that reported on the existing datasets with simulated copies, indicating that partial copy detection in realistic videos remains a challenging problem that deserves future research.From the PDF
Fig. 7 Precision-recall curves of large-scale copy detection on VCDB, using the temporal network method with different numbers of background distraction videos
Fig. 7 Precision-recall curves of large-scale copy detection on VCDB, using the temporal network method with different numbers of background distraction videosFrom the PDF
Fig. 8 Four frame pairs that are difficult to be detected, which contain very severe and complex content variations
Fig. 8 Four frame pairs that are difficult to be detected, which contain very severe and complex content variationsFrom the PDF

Verification

Claims the paper makes

1 Valid · 3 Not valid · 7 Not checked · 38 Own results and interpretations

30 more claims

Create a free account to see all 49 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 10
  • Probably not a substantial claim. 6

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 24 references

Research works printed but not linked: 1

  1. 11. Law-To, J., Chen, L., Joly, A., Laptev, I., Buisson, O., Gouet-Brunet, V., Boujemaa, N., Stentiford, F.: Video copy detection: a comparative study. In: CIVR (2007) · DOI · Paper page

Cited by

See all 83

Metadata

Authors
Jiang, Y.-G., Jiang, Y., & Wang, J.
Affiliations
Fudan University (CN)
Corresponding author
  • Jiang, Yu-Gang
  • Jiang, Yudong
Venue
Lecture Notes in Computer Science
Published
2014
Pages
357-371
ISSN
  • 0302-9743 (ISSN-L)
  • 1611-3349
OpenAlex
W2129795158
Cited by
82
References
23
Citations by year
2026: 6 · 2025: 5 · 2024: 13 · 2023: 11 · 2022: 10 · 2021: 10 · 2019: 9 · 2018: 6 · 2017: 4 · 2016: 5 · 2015: 2 · 2014: 1
FWCI
2.15
Type
conference paper
Language
en
Open access
closed
Licence
www.springer.com/tdm (tdm)
Locations
Lecture notes in computer science (publishedVersion)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Video Analysis and Summarization
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image and Video Retrieval Techniques
Keywords
video copy detection, video databases, benchmark dataset, online video, manual annotation, copyright infringement detection

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: VCDB: A Large-Scale Database for Partial Copy Detection in Videos. https://app.cochat.ai/papers/10.1007/978-3-319-10593-2_24
The paper
Jiang, Y.-G., Jiang, Y., & Wang, J. (2014). VCDB: A Large-Scale Database for Partial Copy Detection in Videos. Lecture Notes in Computer Science. https://doi.org/10.1007/978-3-319-10593-2_24

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.