An experimental study of an object recognition system that learns

摘要

The main objective of this paper is to study the performance of a Knowledge-based Object REcognition system that Learns (KOREL). Other objectives of this paper include, firstly, identifying information about edge-junction objects that is particularly useful in recognizing objects in two-dimensional (2D) images. An edge-junction object is an object that can be recognized by the arrangement of its junctions and by whether each edge is curved or straight. Secondly, this paper presents methods to acquire (learn) this information automatically and to use it in object recognition. Thirdly, whereas most previous systems aimed at identifying one or a few target objects, the research reported here addresses a more challenging problem, namely, recognizing any edge-junction objects known to the system. The number of objects to be recognized might be large, their appearances in the images might be slightly different from those previously encountered, and they might be partially occluded. KOREL represents characteristic views of a three-dimensional (3D) object by models derived from 2D images. Three-dimensional objects in a scene are then recognized by matching a 2D image of the scene against object models. KOREL recognizes an object primarily by abstracting its structure, with representations of less comprehensive structures indexing representations of more comprehensive structures. Exact matching is not required, so occlusion and imperfect data are accommodated.