pVLAD: A discriminative image descriptor for image retrieval

For the past few years, the performance of object recognition and retrieval has been substantially boosted, which is largely attributed to the advent of many effective image descriptors. The most representative examples are the Fisher Vector (FV) and the Vector of Locally Aggregated Descriptors (VLAD). In this paper we focus on the latter. The original VLAD descriptor directly accumulates the sums of residual vectors as the final image representation, which, inevitably, ignores the varying discriminative power of the residual vectors. To address this problem, we propose to employ ℓ_p-norm IDF weighting on it, and obtain a novel image signature which we term pVLAD. The pVLAD descriptor puts more emphasis on the discriminative sub-blocks of VLAD to highlight their contributions to the similarity evaluation by calculating ℓ_p-norm IDF weights for respective sub-VLAD vectors. Meanwhile, ℓ_p-norm IDF weighting is helpful for alleviating burstiness phenomenon of visual codewords, and thereby provide superior performance. We evaluate our pVLAD descriptor on two publicly available benchmark databases, i.e. the INRIA Holidays and the University of Kentucky Benchmark (UKB). The experimental results demonstrate the significant advantages of pVLAD over the original VLAD formulation.