Clustering-integrated Horizontal Pyramid Feature Embedding for In-domain Unsupervised Person Re-identification

Abstract

— A cross-domain and unsupervised Re-Identification (Re-ID) is primarily impacted by the challenges presented by domain shift, intra-class heterogeneity across datasets, and the lack of annotated target pictures. When training and testing take place within the same dataset, these influences are lessened in the in-domain context. In real-time, scalable systems, fine-grained part-based representations like unsupervised horizontal pyramid similarity learning struggle to capture adaptive cluster semantics and integrate generalization. Cluster Integrated Horizontal Pyramid Feature Encoding (CI-HPFE), a deep learning architecture, addresses these issues in this study. Horizontal pyramid-based feature extraction encodes high-resolution local and global body components with key improvements in our model. First, the Pyramid-based Horizontal Part Feature Encoder (PHPFE) splits spatial feature maps into multi-scale horizontal sections (levels 1, 2, 4, 8) and extracts robust, granular embeddings via stripe-wise pooling and processing. Besides feature extraction, a cluster-aware embedding module incorporates unsupervised clustering feedback into representation learning. Identity representations can change with cluster centroids. To achieve alignment and discriminability, our hybrid loss function combines local-part granularity with global identity coherence using triplet margin loss, entropy minimization, and cluster consistency loss. To reduce domain shift, Stripe encoders use a Domain-Adaptive Instance Normalization (DAIN) module to learn domain-conditioned affine statistics. For distributed deployment, a Federated Cluster Synchronization Mechanism (FCSM) synchronizes embedding spaces among edge devices without sharing raw picture data, boosting scalability, privacy, and speed. Benchmark datasets use is Market-1501 and DukeMTMC-ReID, which show considerable performance gains, including 5 – 6% accuracy, ~6% mAP uplift, and enhanced resilience during in-domain assessment. For homogeneous-domain person re-identification systems, the suggested architecture thus strikes a balance between methodological familiarity, unique functional modules, and deployment-level practicality.