Event Location: NSH 1507
Bio: Mark Desnoyer is a Ph.D student in the FRC advised by David Wettergreen. He
holds a B.Eng in Electrical Engineering from Cornell University and a
Masters in Robotics from Carnegie Mellon. He has also worked as a software
engineer for Google and CBS Interactive. His research interests include
computer vision, machine learning, sensing technologies and field robotics.

Abstract: One of the fundamental problems in computer vision is that most computer
vision algorithms are too computationally expensive to use in real-time
systems. Though some speed up approaches have been found for specific
algorithms (e.g. SURF vs. SIFT), there is a more general class of
approaches that keep coming up in the literature and in practice. We’ll
call these approaches visual utility approaches. In a visual utility
approach, we first estimate where the useful visual information will be in
an image and then direct our higher level system to process those regions
that are most likely to be useful. This talk will examine the general
structure of visual utility approaches, identify some characteristics of
problems that make them more like to benefit from visual utility and
provide some guidelines for applying visual utility for a given task. We
will also present a detailed analysis of visual utility systems for
pedestrian detection problems, whose results generalize to object detection
systems in general.