Towards Visual Generalization with Graph Deep Learning

While the world of computer vision is constantly facing new challenges and moves towards more complex learning architectures, new research directions are investigating how to extend existing deep learning methods to new types of data. Geometric Deep Learning tackles the challenge to extend the most successful image-based deep learning methods, such as convolutional neural networks, to 3D shapes and graphs. Starting from the great potential of graph convolutional neural networks, the core idea of this thesis project is to tackle image-based tasks via graph-based deep learning methods. It requires to find a bridge that re-combines back the two worlds of images and graphs mapping the former into the latter. Our intuition is that, by exploiting graph-based models, it would be possible to maintain the semantic information within a picture while loosing all the specific idiosyncrasies which limit visual generalization in traditional image-based models. In the first part we concentrated on image segmentation, comparing different algorithms while maintaining a fixed basic Graph Convolutional Neural Network, known as ECC Network. As a second step we moved our attention to the graph network, modifying its structure in terms of depth, type of layers and learning procedure. With this analysis we were able to get state-of-the-art digit classification for each dataset with a significant reduction (10×) in the original number of pixels. Then, we shifted our attention on more challenging cases where training and test data belong to two different distributions. We leverage on recent domain generalization methods that have shown how combining supervised and self-supervised knowledge can improve generalization.