Self-Supervised Deep Learning via Colorization on 3D Point Clouds for Object Part Segmentation

3D computer vision has grown in popularity in the last few years due to its centrality in many innovative industries such as robotics and autonomous driving, and the application of Deep Learning techniques to the 3D medium has given rise to new challenges. Common applications found in literature that target 3D point clouds range from scene semantic segmentation, to object classification, to object part segmentation. This thesis focuses on the latter - a task that requires the classification of different clusters of the 3D object point cloud into object parts - and aims at exploring the type of impact self-supervised learning techniques can have on such task. Self-supervision has been successfully applied to many Deep Learning applications in computer vision, and it requires the model to perform an additional task (known as pretext or auxiliary task) on unlabeled data. As object colors are often tied to their parts, a self-supervised colorization task (where the model is tasked with predicting the color for each data point in the point cloud) is a promising candidate at driving the model into learning good representations of the data, in the hopes it can perform better on the real task (referred to as downstream task). This thesis takes into consideration two of the most common Deep Learning models found in literature for 3D point clouds processing (PointNet and DGCNN) and explores the impact a colorization self-supervised task can have on improving the performance of 3D object part segmentation tasks.