3D object recognition using perceptual components

We propose a novel representation for appearance-based 3D object recognition. Each 3-D object is represented by the spectral histogram of 2D images at different conditions. The spectral histogram encodes implicitly all the images that are perceptually similar and thus each spectral histogram is called a perceptual component. Given a novel 2D image, the perceptual components of objects are used to deter-mine if the target object is present and which one. Component pruning and filter selection are studied. This representation is applied to the COIL-100 dataset and the experiment results are presented. Comparisons with other methods are also presented.