PROJECT TITLE :

Attention in Attention Networks for Person Retrieval

ABSTRACT:

This study applies explicit mapping in reproducing kernel Hilbert spaces to generate attention values for the input feature map in order to generalize the Attention in Attention (AiA) mechanism that was presented in P. Fang et al., 2019. The AiA mechanism is a model that simulates the capability of constructing interdependencies between local and global characteristics through the interaction of inner and outer attention modules. In addition to a basic AiA module known as linear attention with AiA, two non-linear counterparts known as second-order polynomial attention and Gaussian attention have also been proposed. The purpose of these non-linear counterparts is to use the non-linear properties of the input features explicitly by approximating them with the second-order polynomial kernel and the Gaussian kernel, respectively. The name "Attention in Attention Network" refers to the deep convolutional neural network that was outfitted with the proposed AiA blocks (AiA-Net). The AiA-Net is trained to extract a discriminative pedestrian representation, which combines complementary person appearance and corresponding part features. This representation can then be used to identify pedestrians. Extensive ablation studies have shown that the AiA mechanism works as intended, as does the use of non-linear features that are concealed within the feature map for the purpose of attention design. In addition, our strategy achieves significantly better results than the current state-of-the-art method in comparison to a variety of performance indicators. In addition, with the assistance of the proposed AiA blocks, state-of-the-art performance is also achieved in the video person retrieval task.