A Situated Model of Active Vision

There are two main approaches to active vision for classification. The probabilistic approach employs a belief state concerning the class of the instance and selects actions to reduce uncertainty in the belief state. The situated approach directly learns a mapping from observations to actions (including the classification), driven by the performance on the classification task. One of the advantages of the probabilistic approach is that its action selection takes uncertainty into account. In this paper, we show that the situated approach also takes uncertainty into account in its action selection. To this end, we investigate how a situated gaze-control model without any memory can improve its performance over time in a classification task. We show that it does so by reducing the entropy of the posterior distributions of the possible observations. In this manner, the situated gaze-control model reduces class uncertainty in its observations. In addition, we show that the fixation location of the model serves as a memory of previous observations, which allows the memory-less model to verify conjunctions and disjunctions of facial properties.