Enhancing Human Action Recognition through Temporal Saliency

Images and videos have become ubiquitous in every aspects of life due to the growing digital recording devices. It has encouraged the development of algorithms that can analyze video content and perform human action recognition. This paper investigates the challenging problem of action recognition by outlining a new approach to represent a video sequence. A novel framework is developed to produce informative features for action labeling in a weakly-supervised learning (WSL) approach both during training and testing. Using appearance and motion information, the goal is to identify frame regions that are likely to contain actions. A three-stream convolutional neural network is adopted and improved by proposing a method based on extracting actionness regions. This results in less computation as it is processing only some parts of an RGB frame and also interpret less non-activity related regions, which can mislead the recognition system. We exploit UCF sports dataset as our evaluation benchmark, which is a dataset of realistic sports videos. We will show that our proposed approach could outperform other existing state-of-the art methods.