Global relational attention with a maximum suppression constraint for vehicle re-identification

The goal of vehicle re-identification is to identify the same vehicle from multiple cameras, which is a challenging task. There are many solutions to this problem, among which the self-attention mechanism is very popular. It can capture the long-range dependence in an image, thereby suppressing the irrelevant features. Most of the existing designs are based on isolated pairwise query-key interactions to refine a node. They implicitly mine attention patterns without explicitly modeling node weights. In this paper, we propose a global relational attention mechanism, which makes full use of the global dependence of a node to learn and infer its weight value. Global dependence can measure the importance of nodes more robustly and efficiently. To capture more discriminative features, we propose a maximum suppression constraint to adaptively adjust weight values to expand the range of attention. In addition, we design a pair of effective attention modules based on the proposed attention mechanism, that focus on mining the discriminative features related to vehicle identities from the spatial and channel dimensions. We conduct a large number of experiments on the VeRi-776 and VehicleID datasets, and the experimental results demonstrate the effectiveness of our method.