A Person Re-Identification Scheme Using Local Multiscale Feature Embedding with Dual Pyramids

Abstract

In this paper, we propose a new person re-identification scheme that uses dual pyramids to construct and utilize the local multiscale feature embedding that reflects different sizes and shapes of visual feature elements appearing in various areas of a person image. In the dual pyramids, a scale pyramid reflects the visual feature elements in various sizes and shapes, and a part pyramid selects elements and differently combines them for the feature embedding per each region of the person image. In the experiments, the performance of the cases with and without each pyramid were compared to verify that the proposed scheme has an optimal structure. The state-of-the-art studies known in the field of person re-identification were also compared for accuracy. According to the experimental results, the method proposed in this study showed a maximum of 99.25% Rank-1 accuracy according to the dataset used in the experiments. Based on the same dataset, the accuracy was determined to be about 3.55% higher than the previous studies, which used only person images, and about 1.25% higher than the other studies using additional meta-information besides images of persons.