1 # KCF tracker – parallel and PREM implementations
3 The goal of this project is modify KCF tracker for use in the
4 [HERCULES][1] project, where it will run on NVIDIA TX2 board. To
5 achieve the needed performance we try various ways of parallelization
6 of the algorithm including execution on the GPU. The aim is also to
7 modify the code according to the PRedictable Execution Model (PREM).
9 Stable version of the tracker is available from [CTU server][2],
10 development happens at [Github][3].
12 [1]: http://hercules2020.eu/
13 [2]: http://rtime.felk.cvut.cz/gitweb/hercules2020/kcf.git
14 [3]: https://github.com/Shanigen/kcf
18 The code depends on OpenCV 2.4 (3.0+ for CUDA-based version) library
19 and [CMake][13] is used for building. Depending on the version to be
20 compiled you need to have development packages for [FFTW][4],
21 [CUDA][5] or [OpenMP][6] installed.
23 [4]: http://www.fftw.org/
24 [5]: https://developer.nvidia.com/cuda-downloads
25 [6]: http://www.openmp.org/
26 [13]: https://cmake.org/
30 There are multiple ways how to compile the code.
32 ### Compile all supported versions
35 $ git submodule update --init
39 This will create several `build-*` directories and compile different
40 versions in them. If prerequisites of some builds are missing, the
41 `-k` option ensures that the errors are ignored. This uses [Ninja][8]
42 build system, which is useful when building naively on TX2, because
43 builds with `ninja` are faster (better parallelized) than with `make`.
45 To build only a specific version run `make <version>`. For example,
46 CUDA-based version can be compiled with:
52 [8]: https://ninja-build.org/
57 $ git submodule update --init
62 - Use the just created build directory as "Where to build the
65 - Choose desired build options. Each option has a comment briefly
66 explaining what it does.
67 - Press "Generate" and close the window.
75 $ git submodule update --init
78 $ cmake [options] .. # see the tables below
82 The `cmake` options below allow to select, which version to build.
84 The following table shows how to configure different FFT
87 |Option| Description |
89 | `-DFFT=OpenCV` | Use OpenCV to calculate FFT.|
90 | `-DFFT=fftw` | Use fftw and its `plan_many` and "New-array execute" functions. If `std::async`, OpenMP or cuFFTW is not used the plans will use 2 threads by default.|
91 | `-DFFT=cuFFTW` | Use cuFFTW interface to cuFFT library.|
92 | `-DFFT=cuFFT` | Use cuFFT. This version also uses pure CUDA implementation of `ComplexMat` class and Gaussian correlation.|
94 With all of these FFT version additional options can be added:
96 |Option| Description |
98 | `-DASYNC=ON` | Use C++ `std::async` to run computations for different scales in parallel. This doesn't work with `BIG_BATCH` mode.|
99 | `-DBIG_BATCH=ON` | Concatenate matrices of different scales to one big matrix and perform all computations on this matrix. This mode doesn't work with `OpenCV` FFT.|
100 | `-DOPENMP=ON` | Parallelize certain operation with OpenMP. This can only be used with `OpenCV` or `fftw` FFT implementations. By default it runs computations for differenct scales in parallel. With `-DBIG_BATCH=ON` it parallelizes the feature extraction and the search for maximal response for differenct scales. With `fftw`, Ffftw's plans will execute in parallel.|
101 | `-DCUDA_DEBUG=ON` | This mode adds CUDA error checking for all kernels and CUDA runtime libraries. Only works with `cuFFT` version.|
102 | `-DOpenCV_DIR=/opt/opencv-3.3/share/OpenCV` | Compile against a custom OpenCV version. |
105 ### Compilation for non-TX2 CUDA
107 The CuFFT version is set up to run on NVIDIA Jetson TX2. If you want
108 to run it on different architecture, change the `--gpu-architecture
109 sm_62` NVCC flag in **/src/CMakeLists.txt** to your architecture of
110 NVIDIA GPU. To find what SM variation you architecture has look
113 [9]: http://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/
117 No matter which method is used to compile the code, the results will
120 It operates on an image sequence created according to [VOT 2014
121 methodology][10]. You can find some image sequences in [vot2016
124 The binary can be run as follows:
126 1. `./kcf_vot [options]`
128 The program looks for `groundtruth.txt` or `region.txt` and
129 `images.txt` files in current directory.
131 - `images.txt` contains a list of images to process, each on a
133 - `groundtruth.txt` contains the correct location of the tracked
134 object in each image as four corner points listed clockwise
135 starting from bottom left corner. Only the first line from this
137 - `region.txt` is an alternative way of specifying the location of
138 the object to track via its bounding box (top_left_x, top_left_y,
139 width, height) in the first frame.
141 2. `./kcf_vot [options] <directory>`
143 Looks for `groundtruth.txt` or `region.txt` and `images.txt` files
144 in the given `directory`.
146 3. `./kcf_vot [options] <path/to/region.txt or groundtruth.txt> <path/to/images.txt> [path/to/output.txt]`
148 By default the program generates file `output.txt` containing the
149 bounding boxes of the tracked object in the format "top_left_x,
150 top_left_y, width, height".
152 [10]: http://www.votchallenge.net/
153 [11]: http://www.votchallenge.net/vot2016/dataset.html
157 | Options | Description |
158 | ------- | ----------- |
159 | --visualize, -v[delay_ms] | Visualize the output, optionally with specified delay. If the delay is 0 the program will wait for a key press. |
160 | --output, -o <output.txt> | Specify name of output file. |
161 | --debug, -d | Generate debug output. |
162 | --fit, -f[W[xH]] | Specifies the dimension to which the extracted patch should be scaled. It should be divisible by 4. No dimension is the same as `128x128`, a single dimension `W` will result in patch size of `W`×`W`. |
166 * Vít Karafiát, Michal Sojka
168 Original C++ implementation of KCF tracker was written by Tomas Vojir
169 [here][12] and is reimplementation of algorithm presented in
170 "High-Speed Tracking with Kernelized Correlation Filters" paper [1].
172 [12]: https://github.com/vojirt/kcf/blob/master/README.md
176 [1] João F. Henriques, Rui Caseiro, Pedro Martins, Jorge Batista,
177 “High-Speed Tracking with Kernelized Correlation Filters“, IEEE
178 Transactions on Pattern Analysis and Machine Intelligence, 2015
182 Copyright (c) 2014, Tomáš Vojíř\
183 Copyright (c) 2018, Vít Karafiát\
184 Copyright (c) 2018, Michal Sojka
186 Permission to use, copy, modify, and distribute this software for research
187 purposes is hereby granted, provided that the above copyright notice and
188 this permission notice appear in all copies.
190 THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
191 WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
192 MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
193 ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
194 WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
195 ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
196 OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.