next up previous contents
Next: Computational Neuroscience Up: Active Perception Previous: Volumetric Uniqueness and Gaze

Active Object Recognition

F. Callari, F.P. Ferrie In a typical scenario, an autonomous agent, capable of sensing visual information about its environment, is faced with the problem of recognizing objects from (geometrical) models of them. The sensors are affected by noise, the models are only approximate representations of the surrounding real scene, and both these sources of error cause uncertainty in the recovered models. Since the recognition of the objects is model-based, uncertainty in the models translates into ambiguity in the recogntion: different objects may be equally likely to have produced the sensed data, given the available uncertain information. Yet a decision must be made about which objects are present in the scene and/or what further action must be taken. Building on previous results in the related field of active modeling, we propose a unified Bayesian methodology foractively inferring the objects from the data, and in so doing

Extensive experiments performed have produced encouraging results, showing that the new active recognition technique outerforms a model-only based one, as well as simples ``random walk'' approaches.


next up previous contents
Next: Computational Neuroscience Up: Active Perception Previous: Volumetric Uniqueness and Gaze

Thierry Baron
Mon Apr 7 12:54:24 EDT 1997