Context-rich and partial-aware aggregation network for vehicle re-identification

Vehicle re-identification (Re-ID) aims to identify the same object across nonoverlapping cameras, which has garnered significant attention in recent years. However, there is variability in the images captured by different cameras due to pose, lighting, and occlusion factors, which can pose a challenge to the vehicle Re-ID task. Moreover, many existing methods tend to focus on the overall information of the target image while ignoring the contextual information. To address these issues, we propose a context-rich and partial-aware aggregation network, which integrates a tensor recovery module and part-aware module for object Re-ID. The tensor reconstruction module generates a number of rank-1 tensors, which are employed to capture the contextual features of the object in a high-dimensional space. Concurrently, the part-aware module utilizes image segmentation to emphasize the pertinent local details of the target. These two modules are integrated into a parallel network, which enables the joint extraction of discriminative information from the target image, thereby resulting in superior performance. The proposed method has been demonstrated to be effective across a variety of vehicle types and environments.