Consistency ordinary differential equations network for person re-identification

Person re-identification (Re-ID) aims to retrieve images of a given individual across different camera views. When images contain substantial identity-irrelevant information, the discriminability of person features degrades, leading to reduced recognition accuracy or even failure to recognize individuals. To address this issue, a consistency ordinary differential equations network (CODEN) is proposed, utilizing a dynamical system and incorporating a consistency regularization loss for Re-ID. Specifically, CODEN primarily consists of two key modules: the dynamic feature extraction (DFE) module and the stable trajectory evolution (STE) module. The DFE module parameterizes ordinary differential equations to model the dynamical system and dynamically extracts global saliency information by solving these equations using numerical integrators. The STE module perturbs the initial value of the dynamical system and employs a consistency regularization loss to learn robust feature representations, resisting interference from identity-irrelevant information. Experiments are conducted on three datasets, including Market-1501, DukeMTMC-reID, and CUHK03. The experimental results demonstrate the effectiveness of CODEN in enhancing person Re-ID performance.