Comparing active vision models

Performance on visual tasks such as classification can be enhanced by employing active vision systems. Such systems do not passively receive observations, but have to some extent control over the observations they perceive. There are two general approaches to active vision. The first approach to active vision is a probabilistic approach, in which reducing uncertainty on a part of the world state is the central goal. This uncertainty is modelled by a belief state. The second approach to active vision is a behavioural approach, in which successful behaviour is the central goal. For both approaches, there have been considerable research efforts into designing and studying various active vision models. However, it is not clear how the different existing active vision models relate to each other, and what their relative advantages are. In this report, we identify three main types of active vision models in the probabilistic approach and describe them in a common formal framework. The first type of model selects actions on the basis of the mutual information between actions and classes, and is referred to as the Mutual Information model (MI). The second type of model learns an action policy on the basis of entropy loss in the belief state, the Entropy Loss model (EL). The third type of model bases its action selection on the mode of the belief state, the Mode of Belief state model (MB). In addition, we introduce a fourth type of active vision model that is based on the behavioural approach to active vision, the BeHavioural model (BH). Model BH is identical to EL, except that it learns an action policy that achieves a high performance rather than one that achieves entropy loss in the belief state. We compare the four active vision models empirically on a view-based three-dimensional object classification task. The experimental results give insight into the differences between the models. The overall result is that BH generally outperforms the models EL and MB of the probabilistic approach. In addition, within the probabilistic approach model MI has the best classification performance. Besides revealing performance differences between the active vision models, the experimental results also illustrate properties of the relation between the usefulness of active vision, the number of objects involved in the classification task, and the richness of the visual observations of the models.