Recognizing Compositional Actions in Videos with Temporal Ordering
Document
Description
In some scenarios, true temporal ordering is required to identify the actions occurring in a video. Recently a new synthetic dataset named CATER, was introduced containing 3D objects like sphere, cone, cylinder etc. which undergo simple movements such as slide, pick & place etc. The task defined in the dataset is to identify compositional actions with temporal ordering. In this thesis, a rule-based system and a window-based technique are proposed to identify individual actions (atomic) and multiple actions with temporal ordering (composite) on the CATER dataset. The rule-based system proposed here is a heuristic algorithm that evaluates the magnitude and direction of object movement across frames to determine the atomic action temporal windows and uses these windows to predict the composite actions in the videos. The performance of the rule-based system is validated using the frame-level object coordinates provided in the dataset and it outperforms the performance of the baseline models on the CATER dataset. A window-based training technique is proposed for identifying composite actions in the videos. A pre-trained deep neural network (I3D model) is used as a base network for action recognition. During inference, non-overlapping windows are passed through the I3D network to obtain the atomic action predictions and the predictions are passed through a rule-based system to determine the composite actions. The approach outperforms the state-of-the-art composite action recognition models by 13.37% (mAP 66.47% vs. mAP 53.1%).