Speaker: David Novotny, Research Scientist at Facebook AI Research, London, UK. David was previously a DPhil student in the Visual Geometry Group, University of Oxford in collaboration with NAVER LABS Europe, supervised by Diane Larlus and Prof. Andrea Vedaldi.
Abstract: A learnable component that lifts image pixels into a high-dimensional space is an integral part of any modern image recognition system. In this talk, I will present two deep architectures that achieve improved results predominantly due to a careful design of this pixel-embedding step.
The first part of the talk presents a self-supervised architecture that produces pixel-wise descriptors for establishing image-to-image correspondences. The key ingredient is a novel probabilistic introspection learning scheme which filters out unimportant background samples, allowing the network to selectively represent image pixels that have the potential to result in a correct match.
Next, the task of grouping image pixels belonging to an object is addressed. More specifically, we deal with the instance segmentation problem using a deep convolutional architecture that “colors” image pixels with their instance labels. Identifying the convolutional coloring dilemma, a drawback of standard position-agnostic networks that prevents them from solving this task, we propose a simple correction comprising a novel position-sensitive semi-convolutional operator.
Seminars at NAVER LABS Europe are open to the public but space is limited. Please register here.