Similarity learning in computer vision: applications with structured labels

Jose Rodriguez (Xerox Research)

COMPUTER VISION AND ROBOTICS SERIES

DATE: 2013-11-29
TIME: 14:00:00 - 14:45:00
LOCATION: RSISE Seminar Room, ground floor, building 115, cnr. North and Daley Roads, ANU
CONTACT: JavaScript must be enabled to display this email address.

ABSTRACT:
Traditionally, similarity learning algorithms have been mainly applied in computer vision to enforce that images belonging to the same aclassa should exhibit a high similarity, with applications such as k-NN classification or image search. In this situation, the notion of class is encoded through a binary or categorical label. In this talk, I will describe applications of similarity learning to computer vision problems where the notion of class is encoded by a more complex label. The first and main part of the talk deals with the task of prominent object detection. We formulate detection as a retrieval problem and use supervised features and similarities to encode similar images should have the objects are in a similar location, as opposed to images being visually similar. Note that here aclassa represents an object location (i.e. bounding box). The second part of the talk combines this framework with explicit output embeddings and we learn similarities to explicitly match inputs to outputs. We show that this provides an unconventional method for text recognition ain the wilda (where aclassesa are character sequences). The third part deals with transferring semantic image similarities (where the notion of class is a generic object category) to perform large-scale instance-level retrieval (where the class is the particular object instance).