Prof. Konstantinos G. Derpanis gave an invited talk on June 28th (Friday).

Presenter: Prof. Konstantinos G. Derpanis

Title: Who killed the domain expert?

Abstract:

It has been said that deep learning has eliminated the need for the domain expert. While the advent of deep learning has resulted in a major paradigm shift in computer vision, the domain expert is alive and well. Rather than focusing domain expertise on the design of filters, it has instead been refocused on the design of architectures. In this talk, I will share several projects conducted in my lab and with collaborators where domain knowledge plays a critical role. The first project is concerned with training an optical flow predictor without labeled training data, i.e., an unsupervised approach. Second, I will discuss our work on segmentation-aware convolutional networks (ConvNets). This new architecture integrates segmentation information within a ConvNet. It counteracts the tendency of CovNets to mix information across semantically disparate regions and increases their spatial precision for per-pixel tasks. Finally, I will discuss our work on synthesizing dynamic texture using a two-stream architecture. In the analysis stage, each stream is dedicated to capturing a set of appearance and dynamics-related statistics of a target texture, respectively. The synthesis process generates a novel texture by enforcing the feature statistics of the generated texture to match those of the target texture.

Biography:

Prof. Kosta Derpanis received the Honours Bachelor of Science (BSc) degree in computer science from the University of Toronto, Canada, in 2000, and the MSc (supervisors Prof. John Tsotsos and Prof. Richard Wildes) and PhD (supervisor Prof. Richard Wildes) degrees in computer science from York University, Canada, in 2003 and 2010, respectively. For his dissertation work, he received the Canadian Image Processing and Pattern Recognition Society (CIPPRS) Doctoral Dissertation Award 2010 Honourable Mention. Subsequently, he was a postdoctoral researcher in the GRASP Laboratory at the University of Pennsylvania under the supervision of Prof. Kostas Daniilidis. In 2012, he joined the Department of Computer Science at Ryerson University, Toronto, and is now an associate professor. He currently is a research scientist at the Samsung AI Centre Toronto. His main research field of interest is computer vision with emphasis on motion analysis and human motion understanding, and related aspects in image processing and machine learning.