Multi-dimensional Attention Network for Vehicle Re-identification

Vehicle re-identification (vehicle ReID) is the key point in intelligent transportation systems, which mainly focus on extracting global features by a network of surveillance cameras with non-overlapping fields of view. However, vehicle ReID becomes a more difficult task when taking several aspects into account, e.g. inter-class similarity, intra-class variability, point-of-view variability, and Spatio-temporal uncertainty, which makes feature extraction very difficult. To solve such a problem, we introduce the attention mechanism into the Convolutional Neural Network (CNN) to improve the re-recognition ability of the module. A novel attention based CNN is proposed to calculate the attention weights of multi-channels by crosss-dimensional interaction to strengthen the connection between channel attention and spatial force, which enables the network to better extract the features of vehicles. We evaluate the performance of our network on VeRi-776 and VeRi-Wild datasets currently available. The experimental results demonstrate the effectiveness of our proposed method in vehicle re- recognition tasks.