Probabilistic Object and Viewpoint Models for Active Object Recognition.

For mobile robots to perform certain tasks in human environments, fast and accurate object verification and recognition is essential. Bayesian approaches to active object recognition have proved effective in a number of cases, allowing information across views to be integrated in a principled manner, and permitting a principled approach to data acquisition. Existing approaches however mostly rely on probabilistic models which make simplifying assumptions such as that features may be treated independently and that objects will appear without clutter at test time. We develop a number of probabilistic object and viewpoint models which are explicitly designed to cope with situations in which these assumptions fail, and show these to perform well in a Bayesian active recognition setting using test data in which objects appear in cluttered environments with significant occlusion.