Overview
This project was done in the context of the course 'Machine Learning' at EPFL, given by M. Jaggi and N. Flammarion. It consisted in implementing various convolutional neural network architectures and comparing their performance on a road segmentation task : road / non-road pixel classification for satellite images. CNNs were combined with a deterministic non-trainable convolutional step.
Approach
We use the neural network framework from Keras to build our models. Datasets of satellite images (with labels) were provided beforehand by the course administrators.