A Compact Spatial Feature Representation for Image Classification

In this paper, we propose an alternative framework of spatial pyramid matching (SPM) for describing spatial features, which results in a much lower dimensionality of image representation and yields higher performance compared to SPM. By directly integrating the image division information into the appearance descriptor, the ordinary Bag-of-Words (BoW) model can exploit the multi-resolution spatial information efficiently while avoiding the exponentially increased dimensionality of SPM. We design several spatial descriptors, and show that the new framework overcomes the information redundancy in SPM. Our experimental results on two public image databases demonstrate the superiority of the proposed method.