Learning the hierarchy of human activities in multi-view scenarios

Human activities are inherently complex and variable, making them challenging for deep learning models. Yet, this variability does have an underlying hierarchical structure of related action patterns that naturally emerge from unscripted human activity videos. Exploiting this structure can enhance content understanding. The viewpoint from which actions are observed significantly influences the patterns that models detect. Traditional single-view methods, whether egocentric or exocentric, are limited: egocentric views capture detailed user interactions but are prone to occlusions and narrow fields of view, while exocentric views capture broader contexts but are restricted by fixed camera positions. This thesis proposes integrating these complementary perspectives to uncover patterns hidden by any single viewpoint.