The project is set up in an Anaconda environment. Details about versioning can be found in environment.yml.
The location of the given dataset must be specified in config.py. The base folder where images are generated can also be customized.
After verifying these two parameters, the first step is to run load_data.py. This will generate folders with 32 bit tiff images divided into training and validation images and labels, and test images.
When running train.py. the model is trained with the chosen hyperparameters from config.py, and then predictions are generated for all images from the dataset.
In eval_segmentation.ipynb we compare the IoU value over validation and training dataset while changing the bitmask threshold.
Running visualize.ipynb, we can easily see the input image, ground truth and prediction on one line.
For the test set, the ground truth is empty since it is unknown.