Road Segmentation on satellite imagery
Group: Maxime Schoemans, Pedro de Tavora Santos and Yannick Paul Klose
The code is structed in 5 different components, that are represented by the following python files:
This was the first model implemented by the group and it is heavily based on the code given to us by the Machine Learning course team. This model uses a two-layer convolutional neural network in order to make the predictions.
This file implements our first real optimization of the aforementioned model. It uses a fractal architecture with 4-layer structure in order to make the predictions.
This file implements the final iteration of our models: it creates a u-net architecture with a 9 layer architecture, trains it and saves it in a file, to be later used in the run.py file for the prediction on the test set.
This file implements the Jaccard distance loss that is used to compute the loss that is used to train out u-net model.
This is the file that we use to make the final submission file based on the u-net model.
Additionally, we have 2 folders: one containing the training images (satellite images and groundtruth) and the other containing the test images used to make the submission file.
Since our model has close to 35 million parameters it is impossible to train within a reasonable time on a regular computer without the use of external GPU's. So the first step is to download our model file from the google drive link: https://drive.google.com/open?id=17Zh2KGauj2v3iP-04RdJkguKtvM6O-em . and save it in the same folder as the run.py. After you simply run the run.py file and the output will be the aforementioned submission file.
The load_model function will take up a lot of memory space and it takes some time before the submission file is made.