Experimental Results

Experiments were performed using a 25 object database from [1]. The set of available measurements consisted of 744 optical flow images for each object, gathered uniformly about its viewsphere using four different motion directions for which an eigenspace was constructed using a coarsely sampled subset of these images. Probabilistic models were constructed for all the objects in the database according to equation 1. Each of the available optical flow images was then projected onto the eigenspace and the recognition engine was run on the resulting parameters, assuming a uniform prior distribution over the objects of the database. This yielded a posterior distribution for each view (or rather, each motion). Based on this information, aspect graphs were constructed for all the objects of the database, which were in turn used for gaze planning. This section summarises the results of these experiments.