Event Location: NSH 1507
Bio: Larry Zitnick received the PhD degree in robotics from Carnegie Mellon
University in 2003. His thesis focused on large scale inference
algorithms. Previously, his work centered on stereo vision, including
cooperative and parallel algorithms, as well as the commercial
development of a portable 3D camera. Currently, he is a researcher at
the Interactive Visual Media group at Microsoft Research. He developed
the PhotoDNA technology for removing illegal imagery from the web. His
latest work includes object recognition, and computational photography.

Abstract: There are numerous computer vision algorithms for visual (scene and
object) recognition. However, none of these systems come close to human
capabilities. If we study human responses on similar problems, we could
gain insight into which of three factors (1) learning algorithm (2)
amount of training data and (3) features is critical to humans’ superior
performance. In this talk, I describe our initial progress towards this
goal using a series of human studies and machine experiments. Results
demonstrate that choice of features is the main factor impacting accuracy.

In the second part of the talk, I will briefly describe our latest
research in deblurring images with spatially variant blur kernels. Two
approaches are discussed using blind deconvolution and assisting
hardware. Blind deconvolution accepts as input a single image from which
blur kernels and a sharp image need to be inferred. The hardware
assisted approach uses accelerometers and gyroscopes to constrain the
set of blur kernels used for deconvolution.