Grouping Visual Elements in a Meaningful Way for Human & Computer Vision
When and Where
Research in the past three decades has propelled our understanding of both human and machine vision: We can describe much of the neural mechanisms of human visual perception, and state-of-the-art machine vision can detect objects with high accuracy. Yet research on both human and machine vision lacks a detailed understanding of how visual features are grouped with each other and linked to the rich details of meaning that humans naturally assign to visual information. Such grouping is believed to be essential for the fast, efficient parsing of visual information by decreasing the computational complexity of this process.
The human brain manages to group the primordial soup of visual features that make up a complex scene into meaningful objects. Several heuristic rules, including symmetry, parallelism, proximity, and similarity have been known as key aspects of grouping visual parts into meaningful units since the late 19th / early 20th century Gestalt (German for “form”) psychology. While Gestalt grouping laws have been shown to work empirically, almost nothing is known about why they work and how they are computed and used. I will present initial results and future plans for discovering exploring grouping principles that span computational vision, psychophysics, and neuroimaging.
Having been trained as a physicist and computer scientist, Dirk Bernhardt-Walther earned a Ph.D. in Computation and Neural Systems at the California Institute of Technology in 2006, working with Christof Koch on modeling visual attention and object recognition. After a brief stint with John Tsotsos at York University in Toronto he became a Beckman Postdoctoral Fellow at the Beckman Institute at the University of Illinois at Urbana-Champaign. There he worked with Diane Beck and Fei-Fei Li on natural scene perception and on decoding natural scene categories from fMRI data. From 2010 until 2014, Dr. Bernhardt-Walther was an Assistant Professor of Psychology and from 2012 until 2014 Associate Director of the Center for Cognitive and Brain Sciences at The Ohio State University. In 2014, he moved to the University of Toronto, where he is now Associate Professor in the Department of Psychology. In his work Dr. Bernhardt-Walther aims to decipher the neural mechanisms that underlie the perception of complex real-world scenes and on advancing methods for multivariate analysis of neuroimaging data. Currently he is a Visiting Professor at the Samsung Artificial Intelligence Center in Toronto.