Recognition of object categories using affine kernels

With the growing image collection on the web, classifying images has become an actively explored problem. In this paper we present a novel approach to the classification of images depicting objects in a category using the odd-man-out principle of visual categorization. Specifically, we build a model of an object category by noting discriminative features that are commonly observed across the member images of the class. Appearance changes due to pose, illumination and intra-class variations are modeled using multi-scale affine kernels. The best matching affine kernel for a given query image is found as the one that has the largest overlap of discriminable features that are commonly observed across the class. We show that using the odd-man-out principle of IQ tests not only results in better feature selection but also in more robust object class categorization, in comparison to the state-of-the-art methods on large benchmark image datasets.