Compact codebook generation towards scale-invariance

In this paper, we present a novel visual codebook learning approach towards compactness and scale-invariance for dense patch image encoding. Firstly, each image is described as a bag of orderless gridding local patches, each of which is expressed in three scales. Then a unified objective function is proposed to simultaneously enforce the codebook compactness and select the optimal scale for each local patch, and a convergency provable iterative procedure is utilized for optimization. A direct advantage of the new codebook is that each local patch is essentially described by its best scale, and thus shares certain characteristic of SIFT yet not constrained to any salient point detectors. The experiments on PASCAL 07 dataset validate the effectiveness and efficiency of our proposed method for image classification task.