Gaussian distance coding for image classification

The linear coding methods for image classification work by projecting each local descriptor into the codebook, and making a tradeoff between minimizing the projection error and representation sparseness or locality. In this procedure, it is inevitable to lose some discriminative information which may be very important for image classification. In this paper, we alleviate the information loss in the coding procedure by adding Gaussian distance coding (GDC), aiming to capture the discriminative information lost in the Locality-constrained Linear Coding (LLC). Experiments on the Caltech-101 and Caltech-256 database show that our method outperforms the state-of-the-art performance.