Abstract
Recent advances in deep neural networks have led to significant advances in the automatic understanding of actions in videos. The seminar begins by giving an overview of the algorithms used for video classification, then presents several algorithms for localizing actions in a video in time and space. It shows how " tublets " of actions provide the state of the art for spatio-temporal localization of actions, and why modeling the relationships between objects and humans can improve this performance. A large database of action videos is presented. A weakly supervised algorithm for learning human actions in videos is described. This algorithm significantly reduces the cost of annotations needed to train video classification algorithms.
