SCAN: Spatial and channel attention network for vehicle re-identification

Abstract

Most existing methods on vehicle Re-Identification (ReID) extract global features on vehicles. However, as some vehicles have the same model and color, it is hard to distinguish them only depend on global appearance. Compared with global appearance, some local regions could be more discriminative. Moreover, it is not reasonable to use feature maps with equal channels weights for methods based on Deep Convolutional Neural Network (DCNN), as different channels have different discrimination ability. To automatically discover discriminative regions on vehicles and discriminative channels in networks, we propose a Spatial and Channel Attention Network (SCAN) based on DCNN. Specifically, the attention model contains two branches, i.e., spatial attention branch and channel attention branch, which are embedded after convolutional layers to refine the feature maps. Spatial and channel attention branches adjust the weights of outputs in different positions and different channels to highlight the outputs in discriminative regions and channels, respectively. Then feature maps are refined by our attention model and more discriminative features can be extracted automatically. We jointly train the attention branches and convolutional layers by triplet loss and cross-entropy loss. We evaluate our methods on two large-scale vehicle ReID datasets, i.e., VehicleID and VeRi-776. Extensive evaluations on two datasets show that our methods achieve promising results and outperform the state-of-the-art approaches on VeRi-776.