Active Multi-view Object Recognition and Online Feature Selection

Abstract

Adaptive action selection is crucial for object recognition on robots with limited operating capabilities. It offers much desired flexibility in trading off between the costs of acquiring information and making robust and reliable inference under uncertainty. In this paper, we describe an information-theoretic framework that combines and unifies two common techniques: view planning for resolving ambiguities and occlusions and online feature selection for reducing computational costs. Concretely, our algorithm adaptively chooses two strategies: utilize simple-to-compute features that are the most informative for the recognition task or move to new viewpoints that optimally reduce the expected uncertainties on the identity of the object. Extensive empirical studies have validated the effectiveness of the proposed framework. On a large RGB-D dataset, dynamic feature selection alone reduces the computation time at runtime by five folds, and when combining it with viewpoint selection, we significantly increase the recognition accuracy on average by 8–15% absolute, compared to systems that do not use these two strategies. Lastly, we have also demonstrated successfully the effectiveness of the framework on a quadcopter platform with limited operating time.