Abstract
© 2015 Elsevier B.V. All rights reserved.This paper addresses the problem of combining multi-modal kernels in situations in which object correspondence information is unavailable between modalities, for instance, where missing feature values exist, or when using proprietary databases in multi-modal biometrics. The method thus seeks to recover inter-modality kernel information so as to enable classifiers to be built within a composite embedding space. This is achieved through a principled group-wise identification of objects within differing modal kernel matrices in order to form a composite kernel matrix that retains the full freedom of linear kernel combination existing in multiple kernel learning. The underlying principle is derived from the notion of tomographic reconstruction, which has been applied successfully in conventional pattern recognition. In setting out this method, we aim to improve upon object-correspondence insensitive methods, such as kernel matrix combination via the Cartesian product of object sets to which the method defaults in the case of no discovered pairwise object identifications. We benchmark the method against the augmented kernel method, an order-insensitive approach derived from the direct sum of constituent kernel matrices, and also against straightforward additive kernel combination where the correspondence information is given a priori. We find that the proposed method gives rise to substantial performance improvements.