Finding Object Categories in Cluttered Images Using Minimal Shape Prototypes

摘要

We present an algorithm for recognizing object categories as opposed to specific instances, based on matching prototypical object shapes to gray-level images. The central part of the algorithm is the establishment of correspondence between prototype template and image based on finding qualitative shape invariants in the form of order types of sets of points and lines. A central problem of any matching algorithm like this is the rejection of background and foreground clutter in the image resulting in erroneous matches. By deforming the prototype and iterating the computation of correspondence we reject outliers and improve the quality of the matching. Experimental results in terms of locating examples of specific object classes in real gray-level images are presented. The results demonstrate the robustness of the algorithm and make it an interesting candidate for any categorical recognition system such as database indexing.