NU SEDS Research Seminar “Forming Reliable Feature Correspondences and Distortion-free Graph Embedding with Deep Learning”

The Nazarbayev University School of Engineering and Digital Sciences and National Laboratory Astana invites you to the research seminar “Forming Reliable Feature Correspondences and Distortion-free Graph Embedding with Deep Learning”

Presenter: M. Fatih Demirci, associate professor in the department of computer engineering, TOBB
University of Economics and Technology, Turkey

Abstract: Image classification based on graph embedding has been actively studied in many frameworks. Traditionally, these approaches represent images using handcrafted features, construct graphs where vertices show features and edges encode relations between the features, and then perform graph embedding. In the first part of the talk, professor Demirici will present a novel image classification using a distortion-free graph embedding with deep features. Specifically, after extracting deep features from images, the researchers construct a complete graph such that its vertices represent features and its edges show the distances between the corresponding features. The researchers then perform a distortion-free graph embedding under l∞ to represent the input graph as a set of points in the geometric space. Experimental evaluation of the proposed method including a comparison with the previous graph embedding frameworks demonstrates its effectiveness.

In the second part, the speaker will focus on Finding reliable feature correspondences between a pair of images, which is an important problem receiving much interest in various domains. Given a set of optimal solutions associated with different feature correspondences, our objective is to utilize deep neural networks to select the best optimal solution. Instead of a hand-crafted optimal solution selection as done by the previous work, the researchers propose a novel deep learning-based optimal solution selection. To measure the fitness of our framework, the researchers use an experimental setup where the optimal correspondence selected by our approach is utilized to compute the 3D depth of the scene.

Bio: M. Fatih Demirci is an associate professor in the department of computer engineering, TOBB University of Economics and Technology, Turkey. He worked as an associate professor in the computer science department, Nazarbayev University (Kazakhstan) between 2018-2022 and was a postdoc research scientist at Utrecht University (The Netherlands) between 2006-2007. Dr. Demirci received his MSc. and Ph.D. from computer science department, Drexel University (USA). His research interests lie in computer vision and deep learning. Dr. Demirci has been a PI for several research projects supported by The Scientific and Technological Research Council of Turkey and Nazarbayev University. He has been actively serving as a reviewer/technical program committee member/guest editor/area editor for a number of international conferences and journals.

Zoom link: https://nu-edu-kz.zoom.us/j/91290947516?pwd=d2ZBNTkrU2lMOVpaY3Q5Wnd4Y2RkQT09