DIKT: Direct Incremental Kernel-PCA Tracker

DIKT is a robust holistic visual tracking tool that functions using only the appearance of the target in the initial frame of a given video sequence. It is an adaptive algorithm and updates its stored appearance model online, using newly encountered appearances of the object during tracking.

If this does not load after a short while:
  • Refresh the page
  • Ensure JavaScript is enabled
  • Install the latest Flash Player

The tracker is based on an exact framework for online learning with a special family of indefinite (non-positive definite) kernels. DIKT's kernel defines a reproducing kernel Krein Space, and the kernel's reduced set can be represented without introducing an optimisation problem or a computation of pre-images. Thus, it is efficient and exact.

Video examples of its tracking performance, in comparison to other trackers, are shown above. The workflow is shown in the following graphic. Please refer to the related publication for more detail:

S. Liwicki, S. Zafeiriou, G. Tzimiropoulos, M. Pantic. “Efficient Online Subspace Learning with an Indefinite Kernel for Visual Tracking and Recognition”, IEEE Transaction on Neural Networks and Learning Systems, vol. 23, no. 10, pp. 1624 – 1636, Oct 2012. [pdf | page | poster]



You can download the Matlab implementation of our Direct Incremental KPCA Tracker here:

The video sequences were provided by:

Here are the videos with ground truth1:
1You will need the MATLAB class files in DIKT-2013-04-19.zip.