Learning Adaptive Progressive Representation for Group Re-identification

Abstract

AbstractGroup re-identification (re-id) aims to retrieve a group of people across different surveillance cameras. Due to the change of group layout and membership, group re-id is much more difficult than person re-id. How to address these problems for group re-id is under-explored. In this paper, we propose the Adaptive Progressive group representation Learning Network (APLN) which consists of three innovations: First, we propose a progressive group representation method which fuses individual features together with relation features. Second, we propose a member mask to ignore the impact of changes in the number of members. The member mask is beneficial to getting a more robust group representation from volatile group samples. Third, we propose to use group proxy node as the global representation of the group context graph to obtain precise group context graph information by focusing on more significant individuals. Experimental results demonstrate that our proposed method outperforms the state-of-the-art performance on several group re-id datasets. Compared with the previous methods, the parameters of our model are much fewer and the inference speed is faster.KeywordsGroup re-identificationPerson re-identificationTransformer