Detecting and Analyzing Discontinuities in Computer Vision

点击翻译按钮获取中文摘要

The research in this project will develop robust and effective algorithms for detecting and analyzing meaningful image discontinuities, which are critical cues that play an important part in building general image understanding systems. The project will build on promising preliminary research on multi-flash imaging, which uses active illumination to achieve robust detection and labeling of depth discontinuities in images. These ideas will be generalized by varying various aspects of the illumination parameters and developing methods to handle a much wider variety of imaging conditions. The research will develop a framework for robust multi-view stereo by integrating viewpoint variation with active illumination, and it will extend the concept to develop methods for detecting and analyzing other kinds of meaningful discontinuities (in surface normal, illumination, motion, etc.). To pursue these objectives, extensive data collection and experimentation will be performed with a number of image capture designs and methods for detecting and analyzing discontinuities. The research will enable a variety of both short-term and long-term applications in engineering, medicine, art, and entertainment. Real and synthetic data for experiments will be shared publicly with the research community via the web, as will metrics and software for comparing and evaluating techniques. The educational impacts of the project include the involvement of graduate and undergraduate students and new seminar courses that will be developed related to the research theme. The project will address diversity issues by recruiting undergraduate researchers and short-term graduate students through programs that aim to increase the involvement of underrepresented students in science and engineering, and also by the planned collaborations with colleagues and students in areas outside of Computer Science where women are currently better represented.