QuadNet: Quadruplet loss for multi-view learning in baggage re-identification

Highlights：

• The view-aware attentional local features are learned from the discriminative regions in different views. The local features are fused with global features to enhance the representations of baggage images.

• Random local blur is proposed to handle motion blur which is usually found in baggage images. The multi-task learning of materials is used to improve the discrimination of the QuadNet model.

• The proposed QuadNet model is evaluated extensively on the MVB dataset to demonstrate the effectiveness of QuadNet for baggage ReID. To be compared more fairly, the generalization of QuadNet is evaluated on the Market-1501 and VeRi-776 datasets. It achieves the state-of-the-art performance on all three datasets.

摘要

•A novel quadruplet loss is proposed to solve the multi-view problem in baggage ReID. The quadruplet loss equipped with a multi-view sampling strategy effectively reduces the intra-class distances and increases the inter-class distances. To our knowledge, this is the first work that proposes the quadruplet loss for multi-view learning.•The view-aware attentional local features are learned from the discriminative regions in different views. The local features are fused with global features to enhance the representations of baggage images.•Random local blur is proposed to handle motion blur which is usually found in baggage images. The multi-task learning of materials is used to improve the discrimination of the QuadNet model.•The proposed QuadNet model is evaluated extensively on the MVB dataset to demonstrate the effectiveness of QuadNet for baggage ReID. To be compared more fairly, the generalization of QuadNet is evaluated on the Market-1501 and VeRi-776 datasets. It achieves the state-of-the-art performance on all three datasets.