Reinforcement Learning for Decision Making in Sequential Visual Attention

The innovation of this work is the provision of a system that learns visual encodings of attention patterns and that enables sequential attention for object detection in real world environments. The system embeds the saccadic decision procedure in a cascaded process where visual evidence is probed at the most informative image locations. It is based on the extraction of information theoretic saliency by determining informative local image descriptors that provide selected foci of interest. Both the local information in terms of code book vector responses, and the geometric information in the shift of attention contribute to the recognition state of a Markov decision process. A Q-learner performs then explorative search on useful actions towards salient locations, developing a strategy of useful action sequences being directed in state space towards the optimization of information maximization. The method is evaluated in experiments on real world object recognition and demonstrates efficient performance in outdoor tasks.