Advancing Iterative Quantization Hashing Using Isotropic Prior

摘要

It is prevalent to perform hashing on the basis of the well-known Principal Component Analysis (PCA), e.g.,. Of all those PCA-based methods, Iterative Quantization (ITQ) is probably the most popular one due to its superior performance in terms of retrieval accuracy. However, the optimization problem in ITQ is severely under-deterministic, thereby the quality of the produced hash codes may be depressed. In this paper, we propose a new hashing method, termed Isotropic Iterative Quantization (IITQ), that extends the formulation of ITQ by incorporating properly the isotropic prior proposed by. The optimization problem in IITQ is complicate, non-convex in nature and therefore not easy to solve. We devise a proximal method that can solve problem in a practical fashion. Extensive experiments on two benchmark datasets, CIFAR-10 and 22K-LabelMe, show the superiorities of our IITQ over several existing methods.