Implicit Feature Augmentation for Cloth-Changing Person Re-identification

Implicit Feature Augmentation for Cloth-Changing Person Re-identification

연구 분야: Verification

학회: Chinese Conference on Biometric Recognition

Cloth-changing person re-identification (CC-ReID) aims to match the target pedestrian who might change clothes across different cameras. Despite significant progress in CC-ReID, the task remains challenging due to the scarce cloth-changing samples in existing datasets. To address this, we proposed an Implicit Feature Augmentation Network (IFANet) to mimic data diversity in latent feature space. Particularly, our IFANet consists of three main modules. The global conversion module transfers features based on various fine-grained information, such as texture. The part augmentation module decomposes the features into identity-related and identity-free components and recombines them. Additionally, the identity distribution module is designed to constraint the correct classification of those ambiguous features. Extensive experiments on two widely-used CC-ReID datasets demonstrate the effectiveness of our method.