This project is run using personal notebooks contained in the notebooks
-folder.
The final delivery notebook is contained in final.ipynb
YOLOv5 has been the main architecture used up until this point.
In order to train on the LiDAR videos, they first need to be converted to images (frame by frame).
In the src
-directory there is a script for doing exactly this.
python src/dataset_builder.py [--merge] [--patches]
The --merge
flag is optional. When used it merges the three video channels (ambient, intensity and range) into RGB-images.
The --patches
flag is also optional. When used it splits all videos into 8x 128x128 images
The script can also be run via the notebook.
Please refer to the YOLOv5 documentation for installation and use.
Python scripts has been developed using version Python 3.8, but is expected to work with other versions as well
If you are using pip, run the following at the command-line to install the project dependencies:
pip install -r requirements.txt