Lightweight Multi-Branch Feature Complementary Network for Multi-Modal Object Re-Identification

Multi-modal object re-identification (ReID) is a crucial collaboration technology for identifying specific objects across different data sources, serving as the foundation for tasks like multi-object tracking and collaborative surveillance systems. However, the majority of multi-modal ReID models are constrained by a high number of parameters and significant computational demands, limiting their deployment on edge devices. We propose a Lightweight Multi-Branch Feature Complementary Network (LMCNet) for multi-modal object ReID. LMCNet includes three novel module designs. A Multi-branch Omni-Scale Feature Extraction Network, which can capture multi-modal features effectively. The Attention-Modulated Normalization Integration Module can dynamically fuse features across modalities to enhance collaborative decision-making. The Modality Complementary Module, which can integrate features into modality-specific representations. Experiments on the Market1501-MM, RGBNT201, and MSVR310 datasets demonstrate that LMCNet achieves comparable performance to leading methods while substantially reducing parameters and computational complexity, proving its generalization capability and robustness. Specifically, LMCNet contains approximately 11M parameters, representing a reduction of over 70M compared to the best-performing multi-modal models.