Learning with limited data

In recent years, Deep Learning techniques have demonstrated remarkable achievements across various Computer Vision tasks, frequently surpassing human capabilities. Nevertheless, these data-driven methodologies often demand large volumes of annotated data, necessitating laborious and costly manual annotation procedures. The objective of this thesis is to introduce novel methods designed to mitigate this challenge by harnessing knowledge obtained from diverse domains or tasks, even in the presence of limited annotations. This challenge is commonly known as the Transfer Learning problem. Our exploration will delve into the forefront of Transfer Learning, with a predominant emphasis on the advancement of techniques for Domain Adaptation in diverse computer vision tasks. This research journey begins with a comprehensive investigation into 2D Semantic Segmentation, and we demonstrate how other tasks such as Depth Estimation and Edge Detection can enhance the adaptability of models across different visual domains. Subsequently, the exploration extends to the realm of 3D point cloud classification, where the challenges posed by diverse domain shifts are addressed once again exploiting auxiliary tasks such as shape reconstruction or recent Self-Supervised techniques. The proposed works for 2D Semantic Segmentation and 3D point cloud classification lay the foundation for the development of novel frameworks aimed at tackling the challenging task of multi-modal Domain Adaptation for 3D Semantic Segmentation, where multiple sensors such as RGB cameras and LiDARs are available. Finally, we shed some light on a new exciting and emerging topic which is solving common vision tasks on Neural Fields, which are an emerging paradigm used to represent signals such as images or 3D shapes. We will specifically focus on the 3D scenario, and in the context of Transfer Learning, show for the first time how acting directly on Neural Fields allows the possibility to transfer knowledge among different representations such as from 3D point clouds to meshes.