1 # KCF tracker – parallel and PREM implementations
3 The goal of this project is modify KCF tracker for use in the
4 [HERCULES](http://hercules2020.eu/) project, where it will run on
5 NVIDIA TX2 board. To achieve the needed performance we try various
6 ways of parallelization of the algorithm including execution on the
7 GPU. The aim is also to modify the code according to the PRedictable
8 Execution Model (PREM).
10 Stable version of the tracker is available from [CTU
11 server](http://rtime.felk.cvut.cz/gitweb/hercules2020/kcf.git.),
12 development happens at [Github](https://github.com/Shanigen/kcf.).
16 The code depends on OpenCV 2.4 (3.0+ for CUDA-based version) library
17 and cmake is used for building. Depending on the version to be
18 compiled you have to have [FFTW](http://www.fftw.org/),
19 [CUDA](https://developer.nvidia.com/cuda-downloads) or
20 [OpenMP](http://www.openmp.org/) installed.
22 SSE instructions were used in the original code and these are only
23 supported on x86 architecture. Thanks to the
24 [SSE2NEON](https://github.com/jratcliff63367/sse2neon) code, we now
25 support both ARM and x86 architectures.
29 There are multiple ways how to compile the code.
31 ### Compile all supported versions
34 $ git submodule update --init
38 This will create several `build-*` directories and compile different
39 versions in them. If prerequisites of some builds are missing, the
40 `-k` option ensures that the errors are ignored. This uses
41 [Ninja](https://ninja-build.org/) build system, which is useful when
42 building naively on TX2, because builds with `ninja` are faster
43 (better parallelized) than with `make`.
45 To build only a specific version run `make <version>`, for example:
54 $ git submodule update --init
59 - Use the just created build directory as "Where to build the
62 - Choose desired build options. Each option has a comment briefly
63 explaining what it does.
64 - Press "Generate" and close the window.
72 $ git submodule update --init
78 The `cmake` options below allow to select, which version to build.
80 The following table shows how to configure different FFT
83 |Option| Description |
85 | `-DFFT=OpenCV` | Use OpenCV to calculate FFT.|
86 | `-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.|
87 | `-DFFT=cuFFTW` | Use cuFFTW interface to cuFFT library.|
88 | `-DFFT=cuFFT` | Use cuFFT. This version also uses pure CUDA implementation of `ComplexMat` class and Gaussian correlation.|
90 With all of these FFT version additional options can be added:
92 |Option| Description |
94 | `-DASYNC=ON` | Use C++ `std::async` to run computations for different scales in parallel. This doesn't work with `BIG_BATCH` mode.|
95 | `-DOPENMP=ON` | This option can only be used with CPU versions of the tracker. In normal mode it will run computations for differenct scales in parallel. In the case of the big batch mode it will parallelize the feature extraction and the search for maximal response for differenct scales. If Fftw version is used with big batch mode it will also parallelize Ffftw's plans.|
96 | `-DBIG_BATCH=ON` | Concatenate matrices of different scales to one big matrix and perform all computations on this matrix. This mode doesn't work for OpenCV FFT.|
97 | `-DCUDA_DEBUG=ON` | This mode adds CUDA error checking for all kernels and CUDA runtime libraries. Only works with cuFFT version.|
104 ### Compilation for non-TX2 CUDA
106 The CuFFT version is set up to run on NVIDIA Jetson TX2. If you want
107 to run it on different architecture, change the `--gpu-architecture
108 sm_62` NVCC flag in **/src/CMakeLists.txt** to your architecture of
109 NVIDIA GPU. To find what SM variation you architecture has look
110 [here](http://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/).
114 No matter which method is used to compile the code, the results will
117 It operates on an image sequence created according to [VOT 2014
118 methodology](http://www.votchallenge.net/). You can find some image
119 sequences in [vot2016
120 datatset](http://www.votchallenge.net/vot2016/dataset.html).
122 The binary can be run as follows:
124 1. `./kcf_vot [options]`
126 The program looks for `groundtruth.txt` or `region.txt` and
127 `images.txt` files in current directory.
129 - `images.txt` contains a list of images to process, each on a
131 - `groundtruth.txt` contains the correct location of the tracked
132 object in each image as four corner points listed clockwise
133 starting from bottom left corner. Only the first line from this
135 - `region.txt` is an alternative way of specifying the location of
136 the object to track via its bounding box (top_left_x, top_left_y,
137 width, height) in the first frame.
139 2. `./kcf_vot [options] <directory>`
141 Looks for `groundtruth.txt` or `region.txt` and `images.txt` files
142 in the given `directory`.
144 3. `./kcf_vot [options] <path/to/region.txt or groundtruth.txt> <path/to/images.txt> [path/to/output.txt]`
146 By default the program generates file `output.txt` containing the
147 bounding boxes of the tracked object in the format "top_left_x,
148 top_left_y, width, height".
152 | Options | Description |
153 | ------- | ----------- |
154 | --visualize, -v[delay_ms] | Visualize the output, optionally with specified delay. If the delay is 0 the program will wait for a key press. |
155 | --output, -o <output.txt> | Specify name of output file. |
156 | --debug, -d | Generate debug output. |
157 | --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`. |
161 * Vít Karafiát, Michal Sojka
163 Original C++ implementation of KCF tracker was written by Tomas Vojir
164 [here](https://github.com/vojirt/kcf/blob/master/README.md) and is
165 reimplementation of algorithm presented in "High-Speed Tracking with
166 Kernelized Correlation Filters" paper [1].
170 [1] João F. Henriques, Rui Caseiro, Pedro Martins, Jorge Batista,
171 “High-Speed Tracking with Kernelized Correlation Filters“, IEEE
172 Transactions on Pattern Analysis and Machine Intelligence, 2015
176 Copyright (c) 2014, Tomáš Vojíř\
177 Copyright (c) 2018, Vít Karafiát\
178 Copyright (c) 2018, Michal Sojka
180 Permission to use, copy, modify, and distribute this software for research
181 purposes is hereby granted, provided that the above copyright notice and
182 this permission notice appear in all copies.
184 THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
185 WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
186 MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
187 ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
188 WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
189 ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
190 OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.