Event Location: GHC 4405
Bio: Yong Jae Lee is PhD candidate in the ECE Department at the University of Texas at Austin working under the supervision of Kristen Grauman. His main research interests are computer vision and machine learning. Specifically, he is interested in object recognition and discovery, scene understanding, and activity recognition. His thesis focuses on visual category discovery in images and videos.

Abstract: The standard visual recognition pipeline is placed in a supervised learning setting where annotators provide manually labeled images to train category models. The assumption is a closed-world scenario in which all categories are known. However, in many real world settings, such as a robot navigating an unexplored environment, one cannot prescribe all categories of interest. Instead of placing a strict division between the supervised and unsupervised learning paradigms, a more natural approach would be to recognize familiar objects, but then also discover unfamiliar ones. In this talk, I present approaches to discover visual categories in image collections and videos. I show how to leverage object-level context from familiar categories to produce more coherent category discoveries in unlabeled multi-object images, and how to automatically discover and segment the foreground objects in unannotated videos. My results on benchmark datasets show state-of-the-art results for object category discovery and video segmentation.