Towards open-world video understanding and reasoning

The rapid growth of video data has created a pressing need for intelligent systems that can understand dynamic visual scenes beyond predefined categories and narrowly specified tasks. Real-world videos contain unfamiliar objects, complex interactions, temporal dependencies, and ambiguous events, yet conventional approaches often rely on closed vocabularies, extensive annotations, and limited forms of supervision. These constraints hinder their ability to generalize to open and continuously changing environments. This thesis investigates open-world video understanding and reasoning from an object-centric perspective. It develops learning frameworks that integrate spatial perception, temporal association, visual-language alignment, and contextual reasoning. By connecting object-level representations across frames with semantic knowledge expressed through language, the proposed approaches enable models to identify and track previously unseen objects, interpret flexible natural-language instructions, and reason about object behavior and interactions over time. The thesis also explores data-efficient and interactive learning strategies that reduce dependence on dense annotations while allowing models to adapt to limited supervision and human feedback. Together, these contributions establish a unified perspective on how video systems can move from recognizing fixed visual patterns toward understanding objects and events in a flexible, context-aware manner. The research provides methods and evaluation resources for building scalable video intelligence that generalizes across categories, tasks, and environments. It ultimately advances the development of visual systems that can perceive unfamiliar situations, reason about dynamic content, and align their outputs more closely with human intentions.