Abstract

This paper describes a new technique for statistical 3--D object localization. Local feature vectors are extracted for all image positions, in contrast to segmentation in classical schemes. We define a density function for those features and describe a hierarchical pose estimation scheme for the localization of a single object in a scene with arbitrary background. We show how the global pose search on the starting level of the hierarchy can be computed efficiently. The paper compares different wavelet transformations used for feature extraction. 1 Introduction and Motivation Statistical object modeling without segmentation is motivated by speech understanding. Recognition results were strongly enhanced by the idea of incorporating statistical methods in the recognition task [8] there. Recent publications [2, 9, 5] show, that a statistical framework can also successfully be established in image object recognition. Most of the publications construct statistical models based on segmenta...

Citations