Scene understanding and interaction anticipation from first person vision

Egocentric videos can bring a lot of information about actions performed by humans, which can be beneficial for the analysis of human activity by an external agent such as a robot. This external agent is used as a personal assistant, it should be able to anticipate the user's moves, actions and intentions. Action anticipation from first-person cameras is a challenging task due to the nature of the data processed and the non-scripted, as well as person-specific character of actions. This position paper briefly reviews the current state-of-the-art approaches for action anticipation from egocentric video and proposes ideas on how to improve the quality of predictions.