Spring 2016 Colloquium Computer and Information Sciences, Feb 16, 2016, 11:00AM - 12:00PM, SERC 306
Weakly Supervised Learning for Image and Video Analysis
Dr. Richard Souvenir , University of North Carolina at Charlotte
Many modern computer vision methods rely on large-scale labeled datasets to train supervised algorithms. Due to the proliferation of cameras and ease with which images and videos can be stored and shared, amassing large datasets for a variety of problems is relatively easy. However, obtaining ground truth annotations at a large scale usually involves crowdsourcing or other expensive and/or potentially inaccurate methods. In some cases, we can take advantage of the inherent structure of natural image sets and sidestep the costs of curating large, annotated datasets. In this talk, I will describe problems that can be addressed in weakly supervised settings, where the data labels are missing, noisy, and/or ambiguous. This talk will cover three of our recent methods for image and video analysis: (1) human action recognition in multi-camera networks that leverages ambiguous data for recognizing actions and dynamic viewpoint selection; (2) label correction for large-scale image sets; and (3) semi-supervised multi-output image manifold regression for segmentation and pose estimation. These methods are components of interdisciplinary collaborative projects with architects, ethnographers, and physicians.
Richard Souvenir is an Associate Professor in the Department of Computer Science at The University of North Carolina at Charlotte and directs the Video and Image Analysis Lab. His research interests lie in understanding natural video without developing highly constrained models to improve computer vision tasks, such as segmentation, recognition, and tracking in various domains, including biomedical imaging and human motion analysis. Dr. Souvenir received his D.Sc. from Washington University in St. Louis in 2006.