Abstract
Cloth-Changing Person Re-identification (CC-ReID) aims to match pedestrian images captured from different camera views over long periods. The key challenge is to eliminate the disturbance of clothes and excavate the discriminative identity features. However, recent works mostly focus on body shape or contour sketching, the potential pedestrian features are not fully mining. Inadequate information mining may lead to suboptimal performance. In our framework, we propose a template-based clothes replacement method (TCR), which helps to obtain person images with similar but not exactly the same clothing. By doing so, the clothing interference might be largely eliminated while also retaining the identity information as much as possible. Additionally, we propose two relation learning modules based on Transformer to excavate discriminative relation features that exist within and between images, which are the inner-image relation mining module (INR), and inter-image relation mining module (ITR). INR first extracts body key-point features, and then explores the implicit correlation among key-points. Compared to isolated global and local features, the inner-image relation feature is more inclined to mine discriminative structural features. Based on the learned inner-image relation feature, ITR models the relation between pedestrian images. By exploring the relation between images, some distinct features which distinguish one from others could be learned. Finally, extensive experimental results on two publicly available datasets, PRCC, and LTCC demonstrate the effectiveness of our method.