Dual attention granularity network for vehicle re-identification

Vehicle re-identification (Re-ID) aims to search for a vehicle of interest in a large video corpus captured by different surveillance cameras. The identification process considers both coarse-grained similarity (e.g., vehicle Model/color) and fine-grained similarity (e.g., windshield stickers/decorations) among vehicles. Coarse-grained and fine-grained similarity comparisons usually attend to very different visual regions, which implies that two different attention modules are required to handle different granularity comparisons. In this paper, we propose a dual attention granularity network (DAG-Net) for Vehicle Re-ID. The DAG-Net consists of three main components: (1) A convolutional neural network with a dual-branch structure is proposed as the backbone feature extractor for coarse-grained recognition (i.e., vehicle Model) and fine-grained recognition (i.e., vehicle ID); (2) the self-attention model is added to each branch, which enables the DAG-Net to detect different regions of interest (ROIs) at both coarse-level and fine-level with the assistance of the part-positioning block; (3) finally, we obtain refined regional features of the ROIs from the sub-networks ROIs. As a result, the proposed DAG-Net is able to selectively attend to the most discriminative regions for coarse/fine-grained recognition. We evaluate our method on two Vehicle Re-ID datasets: VeRi-776 and VehicleID. Experiments show that the proposed method can bring substantial performance improvement and achieve state-of-the-art accuracy. In addition, we focus on the different effects of regional features and global features. We conduct experiments to verify it in the PKU dataset and discuss the effectiveness.