Mitigating Batch Normalization bias for single domain generalizable person re-identification

Domain generalizable person re-identification (ReID) poses a significant challenge in real scenarios, as it aims at transferring the knowledge learned from single- or multi-source domains to unseen target domains. Due to the deterministic values of feature statistics (mean and standard deviation) learned from source-domain data, Batch Normalization (BN) exhibits a severe bias towards source domain. This bias is particularly pronounced in single-source domain scenarios, often leading to catastrophic performance degradation. In this paper, a novel Debiasing Batch Normalization (DBN) approach is proposed to alleviate the source-domain bias caused by deterministic values. The DBN is composed of two novel components, a Dynamical Transformed Module (DTM) component and a Gallery to Query Test-time Adaptation (G2QTA) component. Specifically, DTM can flexibly generate the feature statistics relying on the input samples, which is beneficial to adapting to unseen domain and addressing the domain shift. Note that DTM is only embedded in high level stages to replace several BN modules. A novel G2QTA is designed to revise deterministic values, which are not modified by DTM, and adapt to the style of unseen target domain. Extensive experiments demonstrate that our DBN outperform most methods on the task ‘Market1501 → DukeMTMC-ReID’ and ‘DukeMTMC-ReID → Market1501’.

Authors

• Luyifu Chen
• Yanfeng Li Department of Computer Science at School of Informatics, Xiamen University
• Jia Sun Bytedance Inc
• Houjin Chen
• Minjun Wang
• Wanru Peng

Keywords

• Domain generalization
• Person re-identification
• Batch normalization
• Dynamic module
• Test-time adaptation