Learning pairwise image similarities for multi-classification using Kernel Regression Trees

We are often faced with the problem of distinguishing between visually similar objects that share the same general appearance characteristics. As opposed to object categorization, this task is focused on capturing fine image differences in a pose-dependent fashion. Our research addresses this particular family of problems and is centered around the concept of learning from example pairs. Formally, we construct a parameterized visual similarity function optimally separating the pairs of images that depict the objects of the same class or identity from the pairs representing different object classes/identities. It combines various image distances that are quantified by comparing local descriptor responses at the corresponding locations in both paired images. To find the best combinations, we train ensembles of so-called Kernel Regression Trees which model the desired similarity function as a hierarchy of fuzzy decision stumps. The obtained function is then used within a k-NN-like framework to address complex multi-classification problems. Through the experiments with several image datasets we demonstrate the numerous advantages of the proposed approach: high classification accuracy, scalability, ease of interpretation and the independence of the feature representation.