aMLP-ReID: A Vision MLP Architecture Mixed Linear Attention for Person Reid

The Vision Transformer architecture based on Self Attention has been found performing well in feature extracting compared with the CNNs, but there were simply few studies on downstream detecting tasks. The person re-identification task lies on judging whether there is a specific person by extracting image features. There is inevitable loss when convolution kernels are operated with pooling and downsampling with traditional methods based on convolutional network. Therefore, we considered replacing Multihead Attention in Vision Transformer mechanism with a MAL layer based on Vision MLP. The attention mechanism aggregates spatial information. Specifically, the constructed MAL layer explored the MLP structure fitting spatial information. Based on the MLP structure, aMLP-reid mixed linear attention with which was proposed for person re-identification. The improvement was obvious compared with the CNNs on several popular datasets after plenty of experiments. The experiments also demonstrated the effectiveness of Vision MLP architecture for image matching tasks.