Multi-receptive field attention for person re-identification

Person re-identification is a challenging yet meaningful task to match two pedestrian images captured from non-overlapping cameras for public security. Attention schemes have been widely applied to deep learning based person re-identification because they help deep networks dynamically focus on salient regions on person images containing large scale variations and unconstrained auto-detection errors. However, previous attention approaches typically are of a single receptive field, which are difficult to capture rich structural affinities from different scales, harming salient region inferring effect. In this paper, we propose a multi-receptive field attention (MRFA) for person re-identification. MRFA captures multi-scale structural affinities among spatial positions in feature maps via weighted summing over correlations among positions represented with convolutional features of different receptive fields. The multi-scale structural affinities are further applied to infer the importance of different spatial locations. The MRFA is embedded into popular deep architectures (e.g., ResNet and Res2Net) to enhance the feature learning effect for person re-identification. The main contribution of this paper is to extend single receptive field attention to multi-receptive field attention to improve person re-identification effectively. Experiments on three public datasets, i.e., Market-1501, DukeMTMC-reID, and MSMT17, demonstrate that our method is superior to state-of-the-art person re-identification approaches, e.g., on the largest MSMT17 dataset, our method's rank-1 identification rate is 83.9%.