Federated Learning across Heterogeneous Domains, Devices and Models

The widespread deployment of Computer Vision systems in real-world applications has exposed fundamental limitations of traditional centralized learning paradigms, particularly in scenarios characterized by distributed and heterogeneous data sources, sensing platforms, and privacy constraints. Modern vision models are increasingly required to operate across diverse viewpoints, environmental conditions, and acquisition domains, while being trained on data that cannot be centrally collected or shared. Federated Learning (FL) has emerged as a compelling framework to address these challenges by enabling collaborative model optimization directly on distributed clients, without exposing raw data. However, its application to Computer Vision remains non-trivial due to the high dimensionality of visual data, strong non-IID client distributions, heterogeneous computational capabilities, and the growing scale of modern architectures. This thesis investigates Federated Learning as a unifying paradigm for privacy-aware and scalable Computer Vision under realistic deployment conditions. The work progresses from the design of large-scale synthetic and mixed-reality datasets for aerial perception, to federated strategies for heterogeneous autonomous agents, and finally to novel pipelines that integrate Foundation Model (FM)s through efficient and privacy-preserving knowledge transfer. Together, these contributions show how decentralized learning can tackle multi-view and multi-domain diversity while preserving robustness and efficiency. Overall, this work advances the state of the art in Computer Vision by proposing data-centric, algorithmic, and system-level solutions that enable robust learning from decentralized, heterogeneous, and privacy-sensitive data sources, bridging the gap between theoretical federated frameworks and the real-world.