Temporal Commonality Discovery


Temporal Commonality Discovery (TCD) ≡ Given two (or more) videos, discover common events in an unsupervised manner.


Code


  • Usage

    In Matlab

    >> cd tcd_v1.2
    >> addpaths
    >> demoTCD_soft
    >> demoTCD_hard

    • Change the variable pickEx for more examples.

  • File Structure

    Files included in tcd_v1.2.zip

    • addpaths.m: Add required paths to Matlab.
    • demoTCD_soft.m: Demo of TCD using soft-clustering
    • demoTCD_hard.m: Demo of TCD using hard-clustering
    • func/: Folder that contains main functions of TCD
    • featkmeans/: A method for getting the histograms. Other clustering methods can be directly applied.

  • Change Log

    2013-01-06

    • Extend TCD with soft clustering and hard clustering
    • Speedup TCD search using integral image
    • Speedup multiple TCD search using queue elimination

    2012-10-06

    • Add the main implementation of the TCD algorithm
    • Add comparisons to the naive SW approach
    • Add a demo for 8 synthetic examples, including detecting multiple commonalities and an application to video indexing

  • Copyright

    Thanks for using the software

    This software is free for use in research projects. If you publish results obtained using this software, please use this citation. Contributing back bugfixes and improvements is polite and encouraged. If you have any questions, feel free to contact Wen-Sheng Chu.

Introduction

  • Temporal Commonality Discovery (TCD)
    • Unsupervised commonality discovery in images has recently attracted much interest (see an example of co-segmentation).
    • In this study, we investigate a relatively unexplored problem to discover common semantic temporal patterns in videos.

  • Contributions
    • To the best of our knowledge, this is the first work that addresses unsupervised discovery of common events in videos.
    • By interpreting the problem into a 2D search space, we propose a branch and bound (B&B) algorithm to efficiently find the global optimum.
    • We derive tight bounds for classical distances between Temporal Bag of Words (TBoW) of two segments, including L1, intersection and Χ2.
    • The framework is general and can be applied to any feature that has been quantified into histograms.


Publications

    Unsupervised Temporal Commonality Discovery
Wen-Sheng Chu, Feng Zhou and Fernando De la Torre
in ECCV 2012, Firenze, Italy.
[ pdf ][ code ][ poster (6.7MB) ][ slides (12MB) ]
  

Reference

Efficient subwindow search: A B&B framework for object localization
C. Lampert, M. Blaschko, and T. Hofmann,
PAMI, 2009.
[ pdf ]
  
Online discovery and maintenance of time series motifs
A. Mueen and E. Keogh,
SIGKDD, 2010.
[ pdf ]