Data- and model-driven multiresolution processing

We introduce a technique for multiresolution processing which elegantly fits in our framework for visual recognition, described in earlier papers. The input is processed simultaneously at a coarse resolution throughout the image and at finer resolution within a small window (fovea). We introduce an approach for controlling the movement of the high-resolution window which allows for both data- and model-driven selection of fixation points. Three fixation modes have been implemented, one based on large unexplained areas in the data, one on conflicts in the object-model database, and one on a 2D "space filling" algorithm. We argue that this kind of multiresolution processing is not only useful in limiting the computational time, as has been widely recognized, but also can be a deciding factor in making the entire vision problem a tractable and stable one. To demonstrate the approach, we introduce a class of 3D surface textures as a feature for recognition in our system. Surface texture recognition typically requires higher-resolution processing than that required for the extraction of the underlying surface. As examples, surface texture is used to discriminate between a ping-pong ball and a golf ball, and "curve texture" is used to recognize different types of gears. Other experimental results also are included to show the advantages and the implications of our approach. © 1996 Academic Press, Inc.