Department of Computer Engineering
S E M I N A R
Recognition of activities through modeling the relationship between temporal context, objects, and scene
Computer Engineering Department
Daily activities from videos can have high intra-class variance and low inter-class variance due to the way different subjects perform the same action. Traditional action recognition techniques that only rely on visual information can have difficulties to distinguish between such actions. Considering that daily activities are likely to be performed in sequential patterns of activities, and that some actions require certain objects and are performed in certain scenes, we propose to introduce a framework to model the relationships between temporal sequence of activities, objects, scene and motion information of actions in a video. Our first experiment is done on a cooking activities dataset, and we show that by combining temporal sequence and motion information of activities, we can improve the overall accuracy of cooking activity recognition task.
DATE: 25 November, 2013, Monday @ 15:40