Camera-aware Embedding Refinement for unsupervised person re-identification

Abstract

We propose a data-centric approach, Camera-aware Embedding Refinement (CER), to enhance the discriminability of unsupervised person re-identification. CER consists of two components: camera proxy memory and Camera-aware Embedding Generation (CEG). Camera proxy memory is initialized with original embeddings and updated during training using auxiliary embeddings generated by CEG to ensure consistency within the memory. The auxiliary embeddings are created by CEG based on the intrinsic relationships within the original dataset, accounting for image variations caused by different camera perspectives. Specifically, CEG handles scenarios such as images of the same person captured by different cameras, images of different individuals captured by the same camera, and images of different individuals from different cameras. Training the downstream unsupervised Re-ID model with only the auxiliary embeddings significantly improves feature discriminability. Our method focuses on generating auxiliary embeddings and can be adapted for various unsupervised Re-ID models. Extensive experiments show that our approach consistently outperforms state-of-the-art techniques.