FMFI: Transformer based four branches multi-granularity feature integration for person Re-ID

Abstract

Extracting multi-granularity features is a critical challenge in person re-identification (Re-ID). While convolutional neural networks (CNNs) are effectively capture local salient features through convolutional kernels, they struggle to construct globally discriminative representations. In contrast, Transformer networks can model long-range dependencies and establish global contextual relationships among features, making them a powerful tool for multi-granularity feature learning in Re-ID. To comprehensively extract multi-granularity features from various aspects such as clothing attributes, walking postures, and social interactions, we propose FMFI, a four-branch multi-granularity feature extraction and integration method based on Transformer. FMFI employs a four-branch architecture to capture diverse feature representations, enhancing the model’s expressiveness and robustness. Specifically, we introduce the Quadra-Net (QN) structure, which extends from the final layer of a replicated Transformer. By appropriately scaling the branch-wise feature weights and aggregating global tokens from all four branches, FMFI constructs enhanced global feature representations. Furthermore, we design the Refined Global Feature (RGF) module, which refines the initial global features and establishes connections with the newly integrated features, leading to more distinctive and discriminative global representations. Extensive experiments on the Market1501, CUHK03, and MSMT17 Re-ID datasets demonstrate that the proposed FMFI method outperforms most existing Re-ID approaches. Our model significantly enriches feature representations and improves the extraction of multi-granularity features, thereby enhancing person re-identification performance.