Active Object Recognition Integrating Attention and Viewpoint Control

sensor parameters to more effectively solve a particular We present an active object recognition strategy which comvision task. One aspect of active vision is the use of an bines the use of an attention mechanism for focusing the search attention mechanism to decide where in the image to for a 3D object in a 2D image, with a viewpoint control strategy search for a particular object. Template matching schemes for disambiguating recovered object features. The attention which move an object template throughout the image offer mechanism consists of a probabilistic search through a hierarno attention mechanism since all positions in the image chy of predicted feature observations, taking objects into a set are treated equally. However, any recognition scheme that of regions classified according to the shapes of their bounding preprocesses the image to extract some set of features contours. We motivate the use of image regions as a focusprovides a basis for an attention mechanism. Assuming feature and compare their uncertainty in inferring objects with the uncertainty of more commonly used features such as lines that the recovered image features correspond to model or corners. If the features recovered during the attention phase features, object search can be performed at those locations do not provide a unique mapping to the 3D object being in the image where the features are recovered. searched, the probabilistic feature hierarchy can be used to For an attention mechanism to be effective, the features guide the camera to a new viewpoint from where the object must be distinguishing, i.e., have low entropy. If the recovcan be disambiguated. The power of the underlying representaered features are common to every object being searched, tion is its ability to unify these object recognition behaviors they offer little in the way of focusing the search for an within a single framework. We present the approach in detail and evaluate its performance in the context of a project providobject. This is typical in object recognition systems which ing robotic aids for the disabled.  1997 Academic Press match simple image features like corners or zeroes of curvature to model features [30, 19, 39, 26]. Although invariant to viewpoint, there may be an abundance of such features