Compact and discriminative representation of Bag-of-Features

Bag-of-Features (BOF) and Spatial Pyramid Matching (SPM) representation have become well-established in image classification and retrieval, owing to its simplicity and good performance. However, the hyper dimensionality of BOF vectors always faces the difficulty of “curse of dimensionality”, and leads to huge computation and storage complexity. To create a compact and discriminative BOF representations, in this paper we propose a novel unsupervised dimensionality reduction framework for the histogram vectors. First, we build the dissimilarity matrix between every histogram pairs, and then perform Multidimensional Scaling technique to obtain a low-dimensional Euclidean embedding of the original BOF while simultaneously preserving the inherent neighborhood structure. The widely used metrics for measuring dissimilarity, including distance and kernel, are investigated to build the dissimilarity matrix as the input of our dimensionality reduction model. Extensive experiment results show that a very low dimension is sufficient for the learning tasks using BOF or SPM without losing the classification accuracy. Comparatively, the state-of-the-art methods cannot achieve high accuracy on the very low dimension. Furthermore, a compact representation of BOFs can improve the classification accuracy compared with the original BOF. Finally, we also show that our compact representation is effective for image retrieval tasks.