This package includes a C++ class with several tracking methods based on the Kernelized Correlation Filter (KCF) [1, 2].
It also includes an executable to interface with the VOT benchmark.

[1] J. F. Henriques, R. Caseiro, P. Martins, J. Batista,
"High-Speed Tracking with Kernelized Correlation Filters", TPAMI 2015.

[2] J. F. Henriques, R. Caseiro, P. Martins, J. Batista,
"Exploiting the Circulant Structure of Tracking-by-detection with Kernels", ECCV 2012.

Authors: Joao Faro, Christian Bailer, Joao F. Henriques
Contacts: [email protected], [email protected], [email protected]
Institute of Systems and Robotics - University of Coimbra / Department of Augmented Vision DFKI

"KCFC++", command: ./KCF
Description: KCF on HOG features, ported to C++ OpenCV. The original Matlab tracker placed 3rd in VOT 2014.

"KCFLabC++", command: ./KCF lab
Description: KCF on HOG and Lab features, ported to C++ OpenCV. The Lab features are computed by quantizing CIE-Lab colors into 15 centroids, obtained from natural images by k-means.

The CSK tracker [2] is also implemented as a bonus, simply by using raw grayscale as features (the filter becomes single-channel).

There are no external dependencies other than OpenCV 3.0.0+. Tested on freshly installed Ubuntu 14.04 / Ubuntu 16.04.

$ mkdir build
$ cd build
$ cmake -DCMAKE_BUILD_TYPE=Release ..
$ make

The frame images are stored in the directory frames. All frames are named by its index (started from 1). The configurations are stored in the file config.txt:

120,205,151,17,50

The first number is the total amount of frames. The last four numbers indicates the initial bounding box parameters of the first frame.

Notes: If your frames are named like "0001.jpg", "02.jpg" or "0100.jpg" (e.g. datasets from Visual Tracker Benchmark), you should run the script below to normalize them to "1.jpg", "2.jpg" and "100.jpg", respectively:

$ python norm.py

Suppose you've finished the procudures above. You could run the KCF tracker with the commands below:

$ cd build
$ ./KCF show

The bounding box parameters of each frame will be stored to a file named result.txt.

The runtracker.cpp is prepared to be used with the VOT toolkit. The executable "KCF" should be called as:

./KCF [OPTION_1] [OPTION_2] [...]

Options available:

gray - Use raw gray level features as in [1].
hog - Use HOG features as in [2].
lab - Use Lab colorspace features. This option will also enable HOG features by default.
singlescale - Performs single-scale detection, using a variable-size window.
fixed_window - Keep the window size fixed when in single-scale mode (multi-scale always used a fixed window).

To include it in your project, without the VOT toolkit you just need to:

// Create the KCFTracker object with one of the available options
KCFTracker tracker(HOG, FIXEDWINDOW, MULTISCALE, LAB);

// Give the first frame and the position of the object to the tracker
tracker.init( Rect(xMin, yMin, width, height), frame );

// Get the position of the object for the new frame
result = tracker.update(frame);