Introduction

This repository provides C++ implementations for two correlation filter-based trackers. The code implements modified versions of the visual trackers proposed in [1] and [2]:

• KCFcpp: This tracker is a C++ port of the Matlab implementation of the kernelized correlation filter (KCF) tracker proposed in [1]. Project webpage: http://home.isr.uc.pt/~henriques/circulant/ KCFcpp uses as default scale adaption the 1D scale filter proposed in [2]. In addition, a fixed template size, the subpixel/subcell response peak estimation, and the model update from [3] is used as in the KCF version used by Henriques et al. in the VOT challenge 2014 (http://votchallenge.net/vot2014/). The scale adaption used by Henriques et al. in the VOT challenge 2014 is available as option.
• DSSTcpp: This tracker is a C++ port of the Matlab implementation of the discriminative scale space tracker (DSST) proposed in [2]. The default settings use a fixed template size and the subpixel/cell response peak estimation as in the KCF version. Project webpage: http://www.cvl.isy.liu.se/en/research/objrec/visualtracking/scalvistrack/index.html

Both implementations use the FHOG features proposed in [4]. More specifically, the FHOG implementation from [5] is used. Both trackers offer the option to use the target loss detection proposed in [6].

Build

Dependencies

• C++11
• OpenCV 3.0
• CMake
• SSE2-capable CPU

Compilation has been tested on Windows 7 with Visual Studio 2013 Ultimate, on Windows 8.1 with Visual Studio 2013 Community and on Ubuntu 14.04 with g++.

Windows 7

• Set environment variables according to OpenCV Setup - Environment Variables
• Launch cmake-gui, create a build folder and configure.
• Open CfTracking.sln in Visual Studio and compile the projects DSSTcpp and KCFcpp.

Ubuntu 14.04

• Install OpenCV 3.0 and CMake.
• Configure and compile:

mkdir <src-dir>/build
cd <src-dir>/build
cmake ../
make -j 8

Usage

• To track images from a webcam, simply launch DSSTcpp(.exe) or KCFcpp(.exe) and mark an object with a rectangle.
• To pass a predefined bounding box, use the -b x,y,w,h command line switch. Boxes are expected to use images starting at position 0,0.
• To track an image sequence or video, copy the contents of <src-dir>/sample/* to your build/release folder and run the batch/sh file. The example launch scripts are brief and explain the trackers' usage. If you run the tracker from Windows cmd, use only one % sign to specify the naming convention of the image sequence.
• To enable target loss detection, run the tracker with the --para_enable_tracking_loss command line switch.
• To achieve tracking performance as close to the original Matlab implementations as possible, run the trackers with the --original_version command line switch. While the trackers are implemented closely to their original Matlab implementations, implementation differences do still exist (even with the --original_version switch) and the tracking performance of the C++ implementations may deviate from their original Matlab implementations.
• To see a full list of available options, run the trackers with --help command line switch.

Commercial Use (US)

The code using linear correlation filters may be affected by a US patent. If you want to use this code commercially in the US please refer to http://www.cs.colostate.edu/~vision/ocof_toolset_2012/index.php for possible patent claims.

Contributors

Luka Cehovin: Equalize FHOG performance on AMD and Intel CPUs

3rdparty libraries used:

• Piotr's Matlab Toolbox http://vision.ucsd.edu/~pdollar/toolbox/doc/
• OpenCV http://opencv.org/
• tclap http://tclap.sourceforge.net/

References

If you reuse this code for a scientific publication, please cite the related publications (dependent on what parts of the code you reuse):

[1]

@article{henriques2015tracking,
title = {High-Speed Tracking with Kernelized Correlation Filters},
author = {Henriques, J. F. and Caseiro, R. and Martins, P. and Batista, J.},
journal = {Pattern Analysis and Machine Intelligence, IEEE Transactions on},
year = {2015}

[2]

@inproceedings{danelljan2014dsst,
title={Accurate Scale Estimation for Robust Visual Tracking},
author={Danelljan, Martin and H{\"a}ger, Gustav and Khan, Fahad Shahbaz and Felsberg, Michael},
booktitle={Proceedings of the British Machine Vision Conference BMVC},
year={2014}}

[3]

@inproceedings{danelljan2014colorattributes,
title={Adaptive Color Attributes for Real-Time Visual Tracking},
author={Danelljan, Martin and Khan, Fahad Shahbaz and Felsberg, Michael and Weijer, Joost van de},
booktitle={Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2014}}

[4]

@article{lsvm-pami,
title = "Object Detection with Discriminatively Trained Part Based Models",
author = "Felzenszwalb, P. F. and Girshick, R. B. and McAllester, D. and Ramanan, D.",
journal = "IEEE Transactions on Pattern Analysis and Machine Intelligence",
year = "2010", volume = "32", number = "9", pages = "1627--1645"}

[5]

@misc{PMT,
author = {Piotr Doll\'ar},
title = {{P}iotr's {C}omputer {V}ision {M}atlab {T}oolbox ({PMT})},
howpublished = {\url{http://vision.ucsd.edu/~pdollar/toolbox/doc/index.html}}}

[6]

@inproceedings{bolme2010mosse,
author={Bolme, David S. and Beveridge, J. Ross and Draper, Bruce A. and Yui Man Lui},
title={Visual Object Tracking using Adaptive Correlation Filters},
booktitle={Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2010}}