Building Unified Human Descriptors For Multi-Type Activity Recognition.
Ilaria Gori, J.K. Aggarwal, Michael S. Ryoo · arXiv (Cornell University) · 2015
Activity recognition is an important as well as a difficult task in computer vision. In the past years many types of activities -- single actions, two persons interactions or ego-centric activities to name a few -- have been analyzed. Nevertheless, researchers have always treated such types of activities separately. In this paper, we propose a new problem: labeling a complex scene where activities of different types happen in sequence or concurrently. We first present a new unified descriptor, called Relation History Image (RHI), which can be extracted from all the activity types we are interested in. We then propose a new method to recognize the activities and at the same time associate them to the humans who are performing them. Next, we evaluate our approach on a newly recorded dataset which is representative of the problem we are considering. Finally, we show the efficacy of the RHI descriptor on publicly available datasets performing extensive evaluations.