Home

Awesome

KCF tracker in Python

Python implementation of

High-Speed Tracking with Kernelized Correlation Filters<br> J. F. Henriques, R. Caseiro, P. Martins, J. Batista<br> TPAMI 2015

It is translated from KCFcpp (Authors: Joao Faro, Christian Bailer, Joao F. Henriques), a C++ implementation of Kernelized Correlation Filters. Find more references and code of KCF at http://www.robots.ox.ac.uk/~joao/circulant/

Requirements

Actually, I have installed Anaconda(for Python 2.7), and OpenCV 3.1(from opencv.org).

Use

Download the sources and execute

git clone https://github.com/uoip/KCFpy.git
cd KCFpy
python run.py

It will open the default camera of your computer, you can also open a different camera or a video

python run.py 2
python run.py ./test.avi  

Try different options (hog/gray, fixed/flexible window, singlescale/multiscale) of KCF tracker by modifying the arguments in line tracker = kcftracker.KCFTracker(False, True, False) # hog, fixed_window, multiscale in run.py.

Peoblem

I have struggled to make this python implementation as fast as possible, but it's still 2 ~ 3 times slower than its C++ counterpart, furthermore, the use of Numba introduce some unpleasant delay when initializing tracker (NEW: the problem has been solved in KCFnb by using AOT compilation).

NEWER: I write a python wrapper for KCFcpp, see KCFcpp-py-wrapper, so we can benefit from C++'s speed in python now.