Variations on the evidence-based object recognition theme

摘要

The Evidence-Based Object Recognition System of Jain and Hoffman (IEEE Trans. Pattern Analysis Mach. Intell.10(6), 783–802 (1988)) has been extended to include some new view-independent features, a new optimized rule generation procedure based upon Minimum Entropy Clustering and a Neural Network which estimates optimal evidence weights and provides an associated matching procedure. This approach provides an objective definition of the difficulty of an object recognition problem and the procedures and performance of the system are evaluated with four sets of CAD (range) models.