Recognition by Adaptive Subdivision of Transformation Space: practical experiences and comparison with the Hough transform

The RAST algorithm combines aspects of search-based recognition methods and of transformation-space methods such as the Hough transform and others. Recognition of the model in the image consists of finding a transformation (say, a translation) under which many model features match image features well. The Hough transform uses a binning approach in which the space of possible transformations (often called `parameter space´) is divided into buckets. As the correspondences between model and image features are considered, votes are cast for the transformation(s) determined by these correspondences. At the end, the bucket with the largest number of votes is considered to represent the best transformation. Problems limiting the applicability of this approach are addressed with an algorithm which implicitly constructs and evaluates a multiresolution Hough transform whose finest buckets are as small as the recursive step termination rectangles