Reinforcement Learning of 3-d Object Recognition from Appearance

An active observer with the task to identify a three-dimensional object is involved in a search for discriminative viewpoints. This paper deenes the recognition process as a sequential decision problem with the objective to disambiguate initial object hypotheses. Reinforcement learning provides an eecient method to evaluate the action sequences and to develop a sensorimotor mapping for autonomous, near-optimal control. The proposed system learns object models from visual appearance and uses a radial basis function (RBF) network for a probabilistic interpretation of the two-dimensional views. The information gain from temporal fusion of object hypotheses serves as utility measure that eventually drives recognition control to eecient performance.