UC San Diego

Recognizing objects is an essential part of navigating through the visual world. Identifying objects and finding boundaries between them provides us with some of the richest sensory information. Similarly, image segmentation and object recognition are among the most fundamental problems in computer vision and machine intelligence. The potential interaction between these processes has been discussed for many years. The usefulness of recognition for segmentation was demonstrated with various top-down segmentation algorithms; however, the impact of bottom-up image segmentation for object recognition is not well understood. One impeding factor is the unsatisfactory quality of image segmentation algorithms. In this work, we take advantage of a recently proposed method for computing multiple stable segmentations and illustrate the application of bottom-up image segmentation as a preprocessing step for object recognition. In parallel to segmentation, the task of visual object recognition is often greatly facilitated by the objects' surroundings. Contextual information can play the very important role of reducing ambiguity in objects' visual appearance. In this dissertation, we propose a new model for object recognition that incorporates two types of context -- co- occurrence and relative location -- with local appearance- based features, thus named CoLA (for Co-occurrence, Location and Appearance). Since a number of contextual models for recognition have been proposed in the recent history, it is necessary to compare the newly proposed model to the existing ones. Over the years, two general kinds of such models have emerged: those with contextual inference based on the statistical summary of the scene, and models representing the context in terms of relationships among objects in the image. Understanding the theoretical and practical properties of such approaches is essential in designing object recognition systems. We provide an analytical analysis of these models and evaluate them empirically