Stochastic Attributed Relational Graph Matching for Image Near-Duplicate Detection

Attributed Relational Graph (ARG) is a useful model for representing many real-world relational patterns. Computing the similarity of ARGs is a fundamental problem for ARG based modelling. This report presents a novel stochastic framework for computing the similarity of ARGs, which defines the ARG similarity as the likelihood ratio of the stochastic process that transforms one ARG to the other. We show that the transformation likelihood can be factorized and approximately calculated using variational approximation and Loopy Belief Propagation. Furthermore, we show that the similarity measure can be learned from training data using a variational E-M process in an unsupervised manner by using graph-level annotations. We apply the technique to visual scene matching and establish a part-based similarity measure for detecting Near-Duplicate Images in video database.