Video Copy Detection without Feature Extraction

Target: Enable detection of copy/duplicate videos in very large scale video archives.

Summary: State-of-the-art content based copy detection methods achieve quite successful detection rates for certain attacks. However, their heavy computational burden prevents them from being suitable for very large scale video archives. To overcome this problem we utilise Motion Co-occurrence Feature (MCF), which makes use of motion vectors that are already available in the video bitstream. Hence feature extraction is almost diminished. Tests carried out with Trecvid 2009 test set indicates that this method can achieve same performance with the best methods for some of the attacks.

Publications: in preparation


Test Set: TRECVID 2009 set [1]

  • 400 hours of reference video
  • 134 query videos per attack. Temporal duration ranges from 3 sec to 3 minutes. Head and tails may contain out of reference video parts.

T2: Picture in picture Type 1 (The original video is inserted in front of a background video)
T3: Insertions of pattern
T4: Strong reencoding
T5: Change of gamma
T6: Decrease in quality -- This includes choosing randomly 3 transformations from the following: Blur, change of gamma (T4), frame dropping, contrast, compression (T3), ratio, white noise
T8: Post production -- This includes choosing randomly 3 transformations from the following: Crop, Shift, Contrast, caption (text insertion), flip (vertical mirroring), Insertion of pattern(T2), Picture in Picture type 2 (the original video is in the background)
T10: change to randomly choose 1 transformation from each of the 3 main categories.


T3 0.955
T5 0.955
T6 0.776
T8 0.440
T10 0.298

Comprasion with Trecvid 2009 systems [2]

Not suitable for picture in picture (T2) and strong compression (T4), where query is rendered to be useless.

[2] TRECVID 2009--Goals, Tasks, Data, Evaluation Mechanisms and Metrics,