Learning the structure of object categories from incomplete supervision

David Novotny (University of Oxford, UK)

Abstract:

The recent successes of methods that infer the geometric and semantic structure of object categories have been primarily fuelled by the availability of increasingly large quantities of labelled data. However, it is unclear whether manual supervision will be able to keep up with the demands of increasingly sophisticated and data-hungry algorithms. In this talk, I will present approaches that tackle the problem of learning the structure of object categories in a weakly supervised fashion.

In more detail, the first part of the talk describes a novel deep network for learning 3D object categories that is cued solely by observing objects from a moving vantage point. Secondly, a self-supervised architecture that produces pixel-wise descriptors for establishing image-to-image correspondences is presented. The key ingredient is a novel probabilistic introspection learning scheme which filters out unimportant background samples. Finally, the task of grouping image pixels belonging to an object is addressed. More specifically, we deal with the instance segmentation problem using a deep semi-convolutional architecture that “colors” image pixels with their instance labels.

Bio:

David Novotny is a Research Scientist at Facebook AI Research, London, UK. Previously, he was a DPhil student in the VGG group, University of Oxford in collaboration with Naver Labs Europe, supervised by Dr. Diane Larlus. and Prof. Andrea Vedaldi. While working as a researcher at CMP Prague under the supervision of Prof. Jiri Matas, he studied at the Czech Technical University and received his MS degree (with honours) in computer vision and machine learning in 2015. His current research interests are weakly supervised representation learning, matching, single-view 3D reconstruction, pose estimation and instance segmentation.