Compact spatial pyramid representation: The viewpoint of codeword

Abstract

Spatial pyramid model (SPM) has emerged as a popular model to incorporate the spatial information of an image into a single vector named spatial pyramid represen tation (SPR). However, the dimensionality of SPR is huge, incurring high computational cost and storage complexity. To solve this problem, we propose a method for producing compact and discriminative image representation. The fact that the spatial distributions of codewords are both different and discriminative, is taken into account in our method. The novel method is closely related to the discriminative SPM (dSPM), which can be seen as a special case of our method. Experimental results on three public datasets show that our method can lead to superior performance to dSPM and SPM with compact image representation.