3D-Guided Multi-Feature semantic enhancement network for person re-ID

Real-world applications require person re-identification (re-ID) systems to effectively utilize 3D shape information. Such information provides richer spatial, depth, and structural understanding to the model, enhancing its robustness to changes in viewpoint and pose. However, most existing person re-ID methods learn feature representations mainly from 2D images, typically ignoring the actual 3D human structure and spatial relationships. To facilitate better deployment of re-ID systems in real environments, we propose a novel framework called the 3D-Guided Multi-Feature Semantically Enhanced Person Re-ID Network. Firstly, for 3D features, we specially design a hybrid global-local deep spatial feature extractor that enhances the understanding and representation of point cloud details and global information perception through centroid connection and offset attention. Furthermore, to address the semantic mismatch between features of different depths in Convolutional Neural Networks (CNNs) and Transformers, we design a Dual Path Interactive Embedding (DPIE) Module. This module improves the discriminative and correlative properties of features, enhancing the model’s ability to understand and adapt to complex features. Finally, we introduce the novel Symbiotic Attention Module (SAM), which addresses the information asymmetry problem caused by the unidirectional flow of traditional cross-attention. SAM enables bidirectional interaction between features, captures richer complementary information, enhances feature fusion, and establishes a deep connection between global structure and local details. We conducted extensive experiments on several widely used person re-ID datasets, including Market-1501, DukeMTMC-reID, and Occluded-REID, and the results demonstrate that our model outperforms existing methods. Notably, on the Market-1501 dataset, our model achieved 95.9% Rank-1 accuracy and 90.1% mAP.