Model was made as a task in the science club : AGH_RACING
It's necessary to clone yolov5 model to the Object_tracking_model independently. I had some difficulties to push changed yolov5 folder as a repo submodule on the github.
>> git clone https://github.com/ultralytics/yolov5
>> cd yolov5
>> pip install -qr requirements.txt
-
Move file: 'cones.yaml' to .\yolov5\data
-
Prepare your dataset. Data is in the Supervisely format originally, but yolov5 requires YOLOv5 PyTorch TXT format. There are two ways of having it done.
You can preprocess data using files in Repo in the following order:
- parse_annotations.py
- extract_images_to_single_folder.py
- test_parsed_annotations.py
- partition_data.py
- Make use of Roboflow app (https://app.roboflow.com) and apply their data converter (Supervisely JSON -> YOLOv5 PyTorch TXT)
- Run 'train_command.txt' on your preprocessed dataset. In case you didn't have GPU I prepared you a ready result from the calculations in the 'model_results' folder.
- Dataset consists of 24 GB photos of cones in Supervisely format. Training set consists 80% of all photos and testing set consists 20% of all photos.
I added other model made in the Google Colab notebook (with free access to their GPU). The notebook is available at the following link :
-
https://colab.research.google.com/drive/1H1gqPZ8LKiqk0qIbZd6B5OKFJxWCMzTM?usp=sharing
Ready for use dataset is smaller, but the number of epochs is greater.