Aug 18: Andrea Vedaldi: Representations, Structure and Efficiency in the Interpretation of Images

CIS Seminar Series
Date & Time: Thursday August 18, 2011 at 10:00 AM
Location: Clark 110

Speaker:  Dr. Andrea Vedaldi
Research Fellow
University of Oxford, UK

Abstract:
Vision, that is the ability of interpreting images, can extract a significant amount of information about the environment, including the category, pose, and mutual relations of the objects within the field of view of the imaging sensor. This power comes at a price: the staggering variety of natural objects and the complex ways in which they interact with viewpoint, lighting, and clutter, make machine vision extremely challenging. Powerful representations, epitomized by our general viewpoint-invariant features, and novel statistical learning techniques have advanced significantly the state of vision technology and have led to algorithms, such as our state-of-the-art multiple-kernel object detector, that are already applicable to real-world problems, including the semantic indexing of images and videos.

Despite this progress, current models capture only a pale sketch of the image content. Our goal is to build into the next generation of visual models a much deeper and detailed understanding of the structure of the scene, including the interaction between multiple objects, their 3D shape, arbitrary viewpoint, and occlusions. For modeling, we propose structured learning, a framework that extends the statistical efficiency of discriminative learning to the problem of inferring complex structures such as semantic segmentations, including learning from weakly labeled data. For the algorithmic efficiency, we review a technique for the approximation of non-linear support vector machines that we recently introduced, and show how this can be combined with bundle optimization or stochastic gradient methods to train complex models from extremely large datasets.

Bio:
Andrea Vedaldi received his PhD and MSc degrees from the Computer Science Department, University of California at Los Angeles, in 2008 and 2005, and the BSc degree (with honors) from the Information Engineering Department, University of Padua, Italy, in 2005. His research interests include the detection and recognition of visual object categories, visual representations, and large scale machine learning applied to computer vision. He is the recipient of the Oxford Glasstone Research Fellowship and of the “Outstanding Doctor of Philosophy in Computer Science” and “Outstanding Master of Science in Computer Science” awards of University of California at Los Angeles.

Comments are closed.