Person Re-Identification Based on Hybrid Dual-Branch ResNet-Transformer Network
Wenjian Duan, Zhaohui Wang · 2024
Person re-identification is an important branch in the realm of computer vision, which aims to match the images of the same individual captured by non-overlapping cameras based on their characteristics. It has broad applications in security surveillance, intelligent transportation systems, and urban planning. Recently, EfficientFormerV2 has achieved a breakthrough due to the utilization of self-attention and the interconnections of global characteristic. Compared with traditional convolutional neural networks, EfficientFormerV2 excels in extracting global abstract features with its global receptive fields and the modeling ability of long-range dependency. Conversely, convolutional neural networks are more adept at extracting local detailed features. To combine the strengths of both, this paper proposes a hybrid Dual-Branch ResNet-Transformer network to capture both global and local features. It comprises a dual backbone, including EfficientFormerV2 and ResNet-50, a global average pooling layer, and a feature fusion module. Specifically, the global branch-based on EfficientFormerV2 backbone-is utilized to extract the global abstract semantic features from images. The local branch-based on the ResNet-50 backbone-is employed to capture local detailed features. Experimental results on the Market1501 and DukeMTMC-reID datasets validate the effectiveness of the proposed model.