DLAReID: double-layer attention network for object re-identification

Improving feature representations is a crucial task in object re-identification (Re-ID). Enhancement of discriminative features and suppression of irrelevant features have become standard approaches through the use of attentions. Long-range dependencies and more comprehensive image information can be considered by self-attentions. Additionally, the importance of pixels can be calibrated by multi-layer perceptrons. In this paper, Double-Layer Attention network for object re-identification (DLAReID) is proposed. Self-attentions and multi-layer perceptrons are combined in the convolutional neural network, which are the main structures in transformers. The limitations of capturing local dependencies in convolutional neural networks and the excessive parameters in transformers are addressed. 3×3 filters are replaced by the proposed double-layer self-attentions in the bottleneck blocks of the last layer of the network. The ability to capture global information is enhanced by the cooperation of the double-branch self-attention of the outer layer and the attention acting on the relative position encoding module of the inner layer. Good performance is achieved by the proposed method, as demonstrated by comprehensive experiments on multiple datasets.