Wang Qizao ; Qian Xuelin ; Li Bin ; Xue Xiangyang ; Fu Yanwei

Abstract

Cloth-changing person Re-IDentification (Re-ID) is a particularly challengingtask, suffering from two limitations of inferior discriminative features andlimited training samples. Existing methods mainly leverage auxiliaryinformation to facilitate identity-relevant feature learning, includingsoft-biometrics features of shapes or gaits, and additional labels of clothing.However, this information may be unavailable in real-world applications. Inthis paper, we propose a novel FIne-grained Representation and Recomposition(FIRe$^{2}$) framework to tackle both limitations without any auxiliaryannotation or data. Specifically, we first design a Fine-grained Feature Mining(FFM) module to separately cluster images of each person. Images with similarso-called fine-grained attributes (e.g., clothes and viewpoints) are encouragedto cluster together. An attribute-aware classification loss is introduced toperform fine-grained learning based on cluster labels, which are not sharedamong different people, promoting the model to learn identity-relevantfeatures. Furthermore, to take full advantage of fine-grained attributes, wepresent a Fine-grained Attribute Recomposition (FAR) module by recomposingimage features with different attributes in the latent space. It significantlyenhances robust feature learning. Extensive experiments demonstrate thatFIRe$^{2}$ can achieve state-of-the-art performance on five widely-usedcloth-changing person Re-ID benchmarks. The code is available athttps://github.com/QizaoWang/FIRe-CCReID.