Abstract

In recent years there has been a renewed interest in appearance based 3-dimensional object recognition. All possible views of an object define its workspace, known as visual workspace. This workspace is coarsely sampled and projected onto a lower-dimensional space and represented as workspace manifold. The dimensionality reduction is usually done using Karhunen-Loeve transform (KLT), or principal component analysis (PCA). The lower dimensional space has been called the eigenspace. For recognition, the test image is projected likewise onto the eigenspace and its position on the appearance manifold is used for the recognition phase. In object recognition, there are many situations in which features based on only second order statistics are not sufficient. To take into account higher order statistics, Independent Component Analysis (ICA) has been proposed. This thesis presents results on appearance based 3-dimensional object recognition accomplished by using ICA. (Abstract shortened by UMI.)