ISR Seminar: Cornelia Fermuller, "A Bio-inspired Approach to Action Understanding" Wednesday, April 26, 2017
3:00 p.m. 1146 A.V. Williams Bldg.
For More Information:
A Bio-inspired Approach to Action Understanding
Cornelia Fermüller Computer Vision Lab and ARC Lab University of Maryland
Roundtable to follow at 4 pm in 1146 AVW
Abstract Current main-stream approaches to Vision relate symbolic information directly to visual input. However, the vision of active agents employs intermediate representations essential for the perception-action cycle supporting the agent’s actions. I will describe my recent work on the computational implementation of such representations. First, to detect and recognize objects, I have developed so-called mid-level grouping processes, which can serve as the interface between image processing and cognition. They have been implemented as image operators to obtain objects in images and image depth data through attention, segmentation and recognition processes. Second, I will describe studies on hand motions and grasps to characterize actions, segment in time, and predict future actions. These computational modules form the core of our robot vision system architecture. The larger goal of this integrative project is to learn from humans using perception how to perform manipulation actions in order to facilitate action learning for robots. Finally, I will discuss the possible applications of this work in the NACS (Neurosciences and Cognitive Sciences).
Biography Cornelia Fermüller is an associate research scientist at UMIACS, University of Maryland. Her research is in the areas of Computer Vision, Human Vision, and Robotics, and she has published more than 40 journal articles and 120 articles in referred conferences and books. She has studied multiple view geometry and statistics, and her work includes view-invariant texture descriptors, 3D motion and shape estimation, image segmentation, and computational explanations and predictions of optical illusions. Her recent work has focused on two topics: the integration of perception, action and high-level reasoning to develop cognitive robots that can understand and learn human manipulation actions, and motion processing with event-based cameras.