CGV: Small: RUI: Analyzing subspace structure for group level image understanding

点击翻译按钮获取中文摘要

The massive growth in the number of images being generated today has fostered a new direction in computer vision research. Images today are almost never generated independently, rather manifest as collections. Therefore, instead of thinking of images as individual entities, we now need to consider images within a group that may have significant correlational structure. Yet, formalizations of several fundamental image analysis tasks such as image segmentation, for the most part, still consider one image at a time. This project makes the case for exploiting this shared structure for understanding content in images. It develops an efficient framework which permits the segmentation of images at a group level and therefore is applicable to various scenarios in which visual data typically presents itself. The primary objective is to design a comprehensive system that addresses all aspects of this problem -- from preconditioning the input using training data, to providing powerful segmentation models that take the group structure into account, to building dictionaries which can then be used to effectively segment a set of related images. This project provides a vehicle for engaging undergraduates to become immersed in research during the semesters and full-time in summers. Students are exposed to and participate in state-of-the-art research in computer vision which furthers their understanding of these topics and stimulate their intellectual curiosity. This can significantly increase the possibility of these students choosing to pursue higher studies in computer science.