Optimized holistic 3D object recognition based on tolerant contour representations

The authors describe a view based 3D recognition system providing recognition and pose estimation of holistically learnt objects. Recognition is based upon biologically motivated contour representations providing tolerance against minor perspective distortions. According to this tolerance, interpolation between different views is implicitly given, and only a small set of prototypical normalized views must be learnt. An active camera fixating the object eliminates two translatory degrees of freedom, while normalization of views eliminates distance and rotation with respect to the camera-axis. So one has to find a set of normalized prototypical views covering completely the view sphere for recognition purposes. A simple heuristic is able to select an almost optimal set of views, which can further be improved using stochastic optimization methods. During recognition the tolerant, normalized representation of a presented object is compared with the prototypes. The best matching prototypes not only provide hypotheses of the object but also give a first estimation of the object pose. A high resolution pose estimation and a precise verification are also possible, if an additional set of reference vectors for the estimation of the view direction is learnt.