Each model within this repository expects the data files to be under the "data" subfolder. For example, training data for the project was under:
data/train/images
data/train/targets
However, the DataReader class (datareader.py) allows any filepaths to be input, so just ensure that your paths are correct.
The data we used hails from VXiew2's challenge dataset, specifically the train and test sets, linked here.
Use the DataReader class to get any data you need to load. It takes PNG images.
See here for example usage:
dr = DataReader("path/to/images", "path_to_targets")
dr.get_file_lists() # dr.get_file_lists_colab() if you're on Colab or Mac
train, val = dr.get_tf_data()
Models 1 through 3 (under "modelX_training.py") are heavily referenced from Kevin Kibe's tutorial on Deep learning for Image Segmentation with Tensorflow, linked here.. Model 4 lightly referenced Derrick Mwiti for U-Net's structure on Keras and Michael Haltuf for inspiration on F1 scoring as a loss metric.
Models 1 and 3 did not merit their own testing, but models 2 and 4 did. As such, models 2 and 4 have their own "modelX_testing.py" files in which they are tested on XView2's testing dataset. Additionally, model 4 is in its own file (model4.py), while model 2 is simply stored in its training files
Each model has a training and results file, each of which hold miscellaneous data surrounding it to support our final writeup. A copy of each model is also stored in a KERAS file for posterity under its respective "modelX_training" folder.