Event Location: NSH 1507
Bio: Hyun Oh Song is a Ph.D. candidate in Computer Science at UC Berkeley. He is a recipient of five year Ph.D. fellowship from Samsung Lee Kun Hee Scholarship Foundation. His research interest lies at the intersection between computer vision, machine learning and optimization with an application focus in large scale object detection with parsimony.

Abstract: Approximately 85% of internet traffic is estimated to be visual data. Conventional object detection algorithms are not yet suitable to harness this unconstrained, massive visual data because they require laborious bounding box annotations for training and large scale inference is unfeasibly slow due to model complexity. In this talk, I will present two instantiations of model parsimony for object detection. For model learning, I will present one-bit object detection: a framework for training object detectors using only one-bit image level annotations of object presence without any instance level annotations (i.e. bounding boxes). This framework provides approximately 70% relative improvement in localization accuracy (as measured by average precision) over the current state of the art weakly supervised learning methods on standard benchmark datasets. For model inference, I will present sparselet models which significantly reduce model inference complexity by utilizing a shared representation, reconstruction sparsity, and parallelism to enable real-time multiclass object detection with deformable part models at 5Hz with almost no decrease in task performance.