A computer’s ability to accurately classify objects in images is a fundamental research focus in computer vision. Current methods of classification typically adopt a multiclass method, in which a detected object in an image is assigned a label from amongst a set of mutually exclusive labels. However, this does not capture the complexity of objects in the real world; For example, “husky”, “dog”, and “puppy” are not mutually exclusive, but while “husky” is automatically a “dog”, it may or may not be a “puppy”.
In Large-Scale Object Classification using Label Relation Graphs, University of Michigan Assistant Professor Jia Deng along with Google co-authors Nan Ding, Yangqing Jia, Andrea Frome, Kevin Murphy, Samy Bengio, Yuan Li, Hartmut Neven, and Hartwig Adam introduce Hierarchy and Exclusion (HEX) graphs, a new formalism allowing flexible specification of relations between labels applied to objects in images.
Using prior knowledge about semantic label relations, the HEX graph formalism allows a “flexible specification of relations between labels applied to the same object: (1) mutual exclusion (e.g. an object cannot be dog and cat), (2) overlapping (e.g. a husky may or may not be a puppy and vice versa), and (3) subsumption (e.g. all huskies are dogs)”, providing a unified classification framework that generalizes existing models.
This work will be presented at the 2014 European Conference in Computer Vision (ECCV ‘14, http://goo.gl/DSOfeC), hosted in Zurich, Switzerland, September 6-12. To learn more, read the full paper at http://goo.gl/1O6RqT.
Source: Google Research