Active Vision Control Policies for Face Recognition using Deep Reinforcement Learning

Robotic systems are capable of interacting with their environment in order to better sense their surroundings. This key ability of robotic systems is often ignored when developing Deep Learning models, since the later are usually trained using static datasets. This limits the ability of robots to perceive the environment in challenging scenarios. On the other hand, integrating perception and action in tightly coupled systems while operating on-the-edge, holds the credentials for deploying DL-enabled robots in such scenarios; Thus leading to more robust agents that can solve challenging tasks more accurately. In this work, we investigate whether active perception approaches can be employed and integrated into robotic systems in order to improve face recognition accuracy, as well as, study the effect of such an approach on the computational requirements for edge applications. To this end, we propose a DRL-based control approach for training agents that are able to identify task-relevant objects, as well as, issue the appropriate control commands to acquire better results. Through the conducted experimental evaluation, we demonstrate that the proposed method leads to significant improvements in face recognition over the rest of the evaluated approaches by providing accurate control commands.