Modeling objects using interest points, edges, and regions

This dissertation explores the application of different low level image features for the task of object category recognition and segmentation. In particular, it proposes novel fast and efficient models to represent object categories using interest points, edges, and regions. For interest points, a part based generate object template is proposed, which models object categories as a joint occurrence of appearance and location of texture patches around interest points. The model makes novel independence assumptions and proposes variational learning techniques which make learning and inference faster. The model is shown to have good detection and localization performance on standard datasets. This template model is enhanced for automatic object segmentation by incorporating image edge information. A model edge response which encodes the position of object edges with respect to the template is learned from image edges. Object segmentation is obtained by integrating the edge and interest point cues in a random field model. Towards object representation using regions, this thesis first presents a multiscale image segmentation algorithm which represents the image as a hierarchical tree of regions, i.e., a segmentation tree. The algorithm models edges between regions as rapidly varying ramp discontinuities, for which parameters are estimated and used to obtain the final segmentation. The thesis then proposes a model that represents object categories in terms of individual regions and their arrangement in the above segmentation tree. The performance of the model is tested with respect to its localization and segmentation accuracy on standard datasets.