ESCNet: Embedding Space Consistency Network With Scale-Perspective Adaptation for Person Re-Identification

Person re-identification (ReID) has made considerable advancements through extensive research in recent years. However, in practical scenarios, multi-scale and perspective-aware person re-identification becomes critically important because of the variations in the distance between the camera and the person and the differences in the angles of view. This paper proposes an Embedding Space Consistency Network (ESCNet). In this network, we first introduce Scale-Perspective Aware Adaptation (SPAA), which allows the model to actively perceive and adapt to diverse variations in scale and perspective during training. Building on this, to address the distribution shift of global features caused by scale and perspective changes, we design a Feature Consistency Loss (FCLoss) that constrains the distance between multi-scale and multi-perspective contrastive samples in the embedding space. Furthermore, to capture fine-grained local information, we propose Local Feature Alignment Enhancement (LFAE), allowing the model to focus on semantic alignment in local regions of the person. Finally, ESCNet achieves progressive feature learning from local discriminability to global consistency through collaborative optimisation of these modules. Extensive experiments on three benchmark ReID datasets show that our approach outperforms existing methods, achieving superior robustness and effectiveness.