ADA framework for unsupervised domain adaptation person re-identification
Wei Zhang, Peijun Ye, Dihu Chen, Tao Su · Pattern Recognition · 2025
Domain shift remains a critical barrier for generalizing person re-identification (ReID) models across datasets. To address this challenge, we present a sparse self-Attention augmented Domain Adaptation (ADA) framework that learns domain-invariant identity features through three key innovations: (1) Sandwich Attention Primitive (SAP), a novel computational unit designed to boost primitive-level domain adaptation. (2) Sparse self-Attention Augmented Bottleneck block (SAAB block), a hierarchical block integrating SAP to enhance adaptation at the architecture level. (3) Scalable Design, if necessary, SAAB block can be flexibly cascaded to construct task-specific ADA framework. Experiments on three benchmarks validate ADA’s superiority: (1) Achieves state-of-the-art performance across domains (e.g., 16.5 % mAP gain on CUHK03 → Market-1501). (2) Demonstrates consistent generalizability and adaptability.