Multimodal Consistency Co-Assisted Training for Person Re-Identification

Person Re-identification is a popular research area in computer vision, which is mainly used to solve the problem of matching target persons in cross-camera pedestrian image gallery. Traditional methods typically fuse all modal inputs into a fixed feature for subsequent key feature extraction and training, which greatly limits the sensitivity of network learning to independent modal specific information. Meanwhile, the concerns of multiple views on the same data will provide extra regulation. To this end, we propose a Multi Modal Collaborative Fusion Network (MMCF) which consists of three components: Multi Branch Cross Attention (MBCA), Hierarchical Attention Coding (HAC), and Consistency assisted Co training Branch (CCB) MBCA can absorb complementary information from other modalities while retaining unique features of the modality; HAC can embed local key information into global features, thereby improving the expression of pedestrian key features. CCB further improves recognition accuracy through collaborative training between multimodal branches and auxiliary branches. Extensive experiments on RGBNT201 dataset verify the effectiveness of our method and demonstrate superior performance over state of the art pedestrian recognition methods.