Search on dual-space: discretization accuracy-based architecture search for person re-identification

Network architectures automatically generated for person re-identification (re-ID) using neural architecture search (NAS) algorithms exhibit unique advantages. However, existing NAS algorithms are primarily designed to solve the image classification task, and person re-ID, as a sub-problem of image retrieval, differs significantly from classification. To address this issue, we propose a neural architecture search method that leverages dual space to tackle the problem of person re-identification. The neural network discovered through this approach is named DSSNet. In our approach, the dual-space framework comprises two distinct subspaces, each housing a specialized re-ID module dedicated to extracting crucial pedestrian information. By integrating the knowledge of person re-identification into the neural architecture search process, DSSNet achieves superior performance and robustness in re-ID tasks. Moreover, traditional gradient-based NAS methods associate the intensity of operations with continuous architecture parameters during the search process, leading to network degradation. To enhance the accuracy of the search, we propose a novel architecture selection method based on discretization accuracy. The method optimizes the selection of architectures by considering their performance at a discrete precision level. In addition, we introduce a retrieval loss to guide the architecture in learning the similarity or dissimilarity between two pedestrian objects. Our approach significantly improves the accuracy of the search process without the need for human intervention. Extensive experiments demonstrate that our final architecture outperforms state-of-the-art re-ID models on three benchmark datasets, showcasing its superior performance in the re-ID task.