Robust Image Set Classification Using Partial Least Squares

Abstract

Image set classification has recently attracted increasing research interest in the field of visual information processing. Different from previous methods that usually characterize set data distribution explicitly using some parametric or non-parametric models, this paper proposes a simple yet effective Partial Least Squares (PLS) regression based method, which seeks to directly learn the underlying statistical relationship between the distributions of set data and their class memberships. With no assumption on the form of set data distribution, the learned model finally reduces to an efficient linear regression from the data space to the class label space, facilitating robust classification of novel test data. Experiments on face recognition and object categorization have shown that the proposed method is competitive to the state-of-the-arts and also quite robust to the noisy set data in practical applications.