Representations for Learning Scene Layout and Inferring Geometric Context from Images

Charless Fowlkes, Professor & Chancellor’s Fellow, Computer Science, University of California, Irvine

When:
Monday, March 16, 2020
12:30pm – 1:30pm

Where:
Metcalf Trustee Ballroom (9th floor)
1 Silber Way, Boston, MA 02215

Abstract: The content of an image can be understood in many ways, from inferring low-level descriptions of surface shape and material properties, to high-level semantic summaries of objects and their interactions. In the field of computer vision, the problems of visual recognition and estimating three-dimensional (3D) geometry have traditionally been approached with very different computational approaches. However, there are many good arguments as to why these tasks should not be treated independently. Accurate recognition and localization of objects in an image should serve to constrain estimates of 3D shape and motion. Similarly, knowledge of scene geometry should provide strong priors and contextual constraints for object recognition and localization.

Professor Fowlkes will describe work from his group that attempts to leverage these connections in order to: (a) predict the geometry of objects and scenes from single images by learning shapes and layouts associated with familiar appearance, and (b) constrain the recognition of objects and estimation of human pose and their interactions using knowledge of scene geometry. A central theme underlying this work is finding representations for 3D shape and layout that are well matched to powerful convolutional neural network architectures used for 2D visual recognition.

Bio: Charless Fowlkes is a Professor in the Department of Computer Science and director of the Computational Vision Lab at the University of California, Irvine. Prior to joining UC Irvine, he received his PhD in Computer Science from UC Berkeley in 2005 and a BS with honors from Caltech in 2000. His research is in computer vision, machine learning and their application to the biological sciences. Dr. Fowlkes is the recipient of the Helmholtz Prize in 2015 for fundamental contributions to computer vision in the area of image segmentation and grouping, the David Marr Prize in 2009 for his work on contextual models for object recognition, and a National Science Foundation CAREER award. He currently serves on the editorial board of Computer Vision and Image Understanding (CVIU) and IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE-TPAMI).