A refactoring and tested version of DeepSort.

• OpenCV (>= 3.4)
• Eigen3
• Boost
• CUDA (>= 10.0)
• Tensorflow ( = 1.14)

For installing the C++ API, I followed the following guide. I included the repository inside the folder tensorflow_cc, where I modified some file for installing the 1.14 version of tensorflow.

For compiling it, after satysfying the requirements listed in the official repository, execute the following command:

cmake -DTENSORFLOW_STATIC=OFF -DTENSORFLOW_SHARED=ON ..
make
sudo make install

Deep Sort works under Linux environments. I recommend a so-called out of source build which can be achieved by the following command sequence:

• mkdir build
• cd build
• cmake ../
• make -j<number-of-cores+1>

Go to the bin diretory and launch the program with the following commands:

./deep_sort /path/to/the/config/file
e.g.:
./deep_sort ../config/deepsort_config.txt

If you want to use it as library, please add tensorflow_framework as well in the target_link_libraries command as showed in the CMakeLists file.

The param file has to be created like the following one:

[DETECTION_FILE]
/path/to/the/detection/file
[IMAGE_DIRECTORY]
/path/to/the/image/folder
[METRIC_MODEL]
/path/to/the/similarity/model
[FEATURE_MODEL]
/path/to/the/feature/extractor
[ARGS_NN_BUDGET]
100
[ARGS_MAX_COSINE_DISTANCE]
0.2
[DT]
1.0
[MAX_IOU_DISTANCE]
0.9
[MAX_AGE]
30
[N_INIT]
3
[CLASSES]
/path/to/the/class/file
[SHOW_DETECTIONS] # 1 = true 0 = false
1

The detection file is a txt file containing the detections in the following format:

#frame_number,#top_left_x,#top_left_y,#width,#height,#class,#confidence
e.g. 0,623,281,16,15,1,0.294832

The class file has to include the names of all the classes detected by your detector.

E.g.:

class1
class2
class3
...
classN

Please, for any comment send me an email to andrea[dot]pennisi[at]gmail[dot]com.