Region to Global Vision Transformer for Baggage Re-Identification

Baggage re-identification(ReID) is a challenging and important task in airport intelligent surveillance, facing the problems of occlusion, background changes, pose differences and highly similar baggage appearance in complex scenes that are difficult to solve in current related research. In order to effectively improve the performance of baggage re-identification, a region-to-global Transformer network(RGViT) is proposed. Three improvement modules including regional self-attention module, global spatial-relation attention module and quadruplet loss function are used to enhance the re-identification performance of the Vision Transformer, which has made outstanding achievements in the field of image processing in recent years. The regional self-attention module further condenses local features into regional features to intensify feature extraction while reducing memory complexity. The global spatial-relation attention module captures the spatial structure of relationships between regional features, stacking them according to their spatial importance distribution to form distinct global contextual information. The quadruplet loss function reduces the intra-class distance of the same baggage identity under different views and increases the inter-class distance, effectively solving the problem of re-identification of high similarity baggage. The proposed method achieves 88.7% Rank-1 accuracy and 89.6% mAP on the baggage re-identification dataset MVB, which has remarkably excellent performance compared with other state-of-the-art methods.