The paper, entitled, "Large-Scale Object Classification using Label Relation Graphs," was co-authored with colleagues from Google, where Prof. Deng has been conducting research for the past year. It addresses a computer's ability to accurately classify objects in images, which is a fundamental challenge in computer vision research and an important building block for many other tasks such as localization, detection, and scene parsing.
Because current classification methods do not adequately capture the complexity of semantic labels in the real world, the authors set out to describe an approach to object classification that exploits the rich structure of real world labels, with a goal of developing a new classification model that allows flexible encoding of relations based on prior knowledge, thus overcoming the limitations of current systems which tend to be either overly restrictive or overly relaxed.
The authors introduce Hierarchy and Exclusion (HEX) graphs, a new formalism that captures semantic relations between any two labels applied to the same object: mutual exclusion, overlap and subsumption. They then provide rigorous theoretical analysis that illustrates properties of HEX graphs such as consistency, equivalence, and computational implications of the graph structure. Next, they propose a probabilistic classification model based on HEX graphs and show that it enjoys a number of desirable properties. Finally, the researchers evaluate their method using a large-scale benchmark and show empirical results which demonstrate that their model can significantly improve object classification by exploiting the label relations.
The researchers' model replaces traditional classifiers such as softmax or independent logistic regressions (left). It takes as input image features (e.g. from an underlying deep neural network) and outputs probabilities consistent with pre-specified label relations.
In 2013, Prof. Deng was awarded the Marr Prize, which is the best paper award of the International Conference on Computer Vision, for additional work in the area of image classification.
Prof. Deng received his Ph.D. from Princeton University in 2012 and his B.Eng. from Tsinghua University, both in computer science. He joined the faculty at Michigan in September 2013. Prof. Deng has built datasets and tools used by more than 1,000 researchers worldwide and his work has appeared in popular press such as The New York Times and MIT Technology Review. He has been co-organizing the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) since 2010. He was also the lead organizer of the BigVision workshops at NIPS 2012 and CVPR 2014.
Posted: September 11, 2014