SFA-Net: Scatter-Fusion Attention Network for Person Re-identification

Abstract The utilization of attention mechanisms has demonstrated promising outcomes in diverse vision tasks, like the Person re-identification (re-ID) task. In the context of person re-ID, the acquisition of effective convolutional features through attention mechanisms plays a crucial role. However, person re-ID presents a unique challenge as individuals manifest themselves through multiple body regions simultaneously. Consequently, a holistic approach is necessary to encode high-order interactions among local features for accurate recognition. To address these challenges, we propose the Scatter-Fusion Attention Network (SFA-Net) as a viable solution. The scatter operation involves the instantiation of multiple attention blocks, enabling simultaneous attention across multiple body areas and the creation of attention maps for these regions. Subsequently, the Fusion operation redistributes these attentions to multiple locations before merging the feature maps into a comprehensive one. SFA-Net effectively leverages the attention mechanism to scatter and fuse attention blocks, facilitating the flow of information among them. The SFA mechanism incorporates spatial and channel-wise attention mechanisms, allowing for a comprehensive emphasis on foreground objects and effective elimination of background clutter, in contrast to conventional attention learning that considers only current features at a time. Extensive experimental results verified that SFA-Net achives state-of-the-art person re-ID on two datasets(Market-1501 and DukeMTMC-reID).