An adversarial framework for open-set human action recognition using skeleton data

Human action recognition is a fundamental problem which is applied in various domains, and it is widelystudied in the literature. Majority of the studies model action recognition as a closed-set problem. However, in real-life applications it usually arises as an open-set problem where a set of actions are not available during training butare introduced to the system during testing. In this study, we propose an open-set action recognition system, humanaction recognition and novel action detection system (HARNAD), which consists of two stages and uses only 3D skeletoninformation. In the first stage, HARNAD recognizes a given action and in the second stage it decides whether theaction really belongs to one of the a priori known classes or if it is a novel action. We evaluate the performance of thesystem experimentally both in terms of recognition and novelty detection. We also compare the system performance withstate-of-the-art open-set recognition methods. Our experiments show that HARNAD is compatible with state-of-the-art methods in novelty detection, while it is superior to those methods in recognition.