Egocentric Action Anticipation

In this PhD thesis, we examine the problem of action anticipation from egocentric (first-person) videos. Action anticipation is the ability to predict and prepare for future events, and is a key component of both human and machine intelligence. Egocentric videos provide a unique perspective on a person's hands and the objects they manipulate, which is crucial for understanding current activities and anticipating the future. We focus on the task of predicting the future activities of the camera wearer, as this has many practical applications. We discuss common challenges in the field of action anticipation from first-person videos and demonstrate the potential of egocentric action anticipation in the domain of personal assistance and health. We also study the problem of untrimmed action anticipation, which involves predicting future actions in real-time as the video is being recorded. Overall, our work explores the field of action anticipation from egocentric videos from different perspectives and has the potential to enable a range of applications in robotics and personal assistance applications.