Understanding human actions through interaction modelling

This research focuses on understanding actions by modelling the interactions between either human & human or human & objects. This is achieved by learning discriminative spatio-temporal representations for recognising human actions at the coarse and finer levels. However, the current state-of-the-art methods are not effective in distinguishing highly similar actions. Such actions that are visually similar to each other and involve subtle interactions between human & objects are known as fine-grained actions. The difficulty in modelling these interactions due to high inter-class similarity, presence of diverse objects, actor and view-point variations have motivated us to propose novel approaches for efficient classification of actions.