Studies in Using Image Segmentation to Improve Object Recognition

Abstract

Recognizing object classes is a central problem in computer vision, and recently there has been renewed interest in also precisely localizing objects with pixel-accurate masks. Since classes of deformable objects can take a very large number of shapes in any given image, a requirement for recognizing and generating masks for such objects is a method for reducing the number of pixel sets which need to be examined. One method for proposing accurate spatial support for objects and features is data-driven pixel grouping through unsupervised image segmentation. The goals of this thesis are to define and address the issues associated with incorporating image segmentation into an object recognition framework. The first part of this thesis examines the nature of image segmentation and the implications for an object recognition system. We develop a scheme for comparing and evaluating image segmentation algorithms which includes the definition of criteria that an algorithm must satisfy to be a useful black box, experiments for evaluating these criteria, and a measure of automatic segmentation correctness versus human image labeling. This evaluation scheme is used to perform experiments with popular segmentation algorithms, the results of which motivate our work in the remainder of this thesis. The second part of this thesis explores approaches to incorporating the regions generated by unsupervised image segmentation into an object recognition framework. Influenced by our experiments with segmentation, we propose principled methods for describing such regions. Given the instability inherent in image segmentation, we experiment with increasing robustness by integrating the information from multiple segmentations. Finally, we examine the possibility of learning explicit spatial relationships between regions. The efficacy of these techniques is demonstrated on a number of challenging data sets.