Diffusion and text-based guidance for image morphing

The School of EECS is hosting the following Visitor Seminar by one of the pioneers in the field of computer vision.

Diffusion and text-based guidance for image morphing

Speaker: Prof Richard Hartley, Australian National University
Host: Dr. Xin Yu

Abstract: The talk will give some background on diffusion as a Generative Model for images, and text-based imaging.  We explore the use of diffusion for estimating probability distributions for natural images.  Based on some theoretical background, we developed methods for finding paths through high probability regions of image space to give natural transitions between different images. The use of text-guidance, and simultaneous geodesic interpolation in image and text space gives smooth and realistic (at least plausible) image morphs.

Bio: Richard Hartley is a Distinguished Professor Emeritus at the Australian National University, where he has worked since 2001.  He obtained a BSc degree in mathematics from the Australian National University followed by a PhD (also in mathematics) at the University of Toronto, and a MSc in Computer Science / Computer Engineering at Stanford University, followed by postdoctoral positions at the University of Frankfurt, Columbia University, Melbourne University and the University of Wollongong.

He worked at General Electric’s Research and Development Centre in Schenectady, NY between 1984 and 2001, where his carried out research in the fields of VLSI Computer Aided Design, and then Computer Vision.  He was also the original Program Leader for, and later team member of the Autonomous Systems and Sensor Technology Program at National ICT Australia (later Data61, part of CSIRO).

He is best known for his research in Computer Vision, particularly geometric and mathematical methods.  This work started with his collaboration with GE's Simulation and Control Systems Division. His research in applying projective geometry techniques to reconstruction using calibrated and semi-calibrated cameras contributed to a dominant theme in computer vision research throughout the 1990s and early 2000s.  In 2000, he co-authored a seminal book, “Multiple View Geometry in Computer Vision”, with Andrew Zisserman, published by Cambridge University Press.  This book summarized a decade's worth of research in this field.

Since taking emeritus status at ANU, he has worked part time for Google Research, Switzerland, as well as various small companies throughout the world.