Foveated Manifold Sensing for object recognition

Abstract

We present a novel method, Foveated Manifold Sensing, for the adaptive and efficient sensing of the visual world. The method is based on algorithms that learn manifolds of increasing but low dimensionality for representative data. As opposed to Manifold Sensing, the new foveated version senses only the most salient areas of a scene. This leads to an efficient sensing strategy that requires only a small number of sensing actions. The method is adaptive because during the sensing process, every new sensing action depends on the previously acquired sensing values. Finally, we propose a hybrid sensing scheme that starts with Manifold Sensing and proceeds with Foveated Manifold Sensing. This sensing scheme needs even less sensing actions for the considered recognition tasks. We apply the proposed algorithms to object recognition on the UMIST and ALOI datasets. We show that, for both databases, we reach a 100% recognition rate with only 10 sensing values.