trainer is based on RetinaNet and detectron2 https://detectron2.readthedocs.io/tutorials/install.html#common-installation-issues
App is written in C++. I recommend QtCreator open-source, and msvc++ from Microsoft Visual Studio Community Edition.
QtCreator can open the cmake file CMakeLists.txt
You need OpenCV, see https://medium.com/beesightsoft/build-opencv-opencv-contrib-on-windows-2e3b1ca96955
Please use OpenCV 3.4, it has YOLOv4 support.
Collect your own dataset and put it into this directory, call the folder data. The structure is described in the report.
app consists of several projects
OBJECT_DETECTOR is the important project, it loads the model found in the data to count cod and saithe.
DATASET_TOOL is for improving the dataset, to create new labels
See Releases for executables of the project. You will find Windows and MacOS binaries. The Windows binaries include pre-compiled OpenCV libraries.
train.py will train a model with RetinaNet
inference.py will create a video with the RetinaNet model
report consists of my bachelor thesis