Lifelong vehicle re-identification via ResNet50 with depthwise spatial attention
Abstract
The core task of vehicle re-identification is to recognize vehicles with the same identity across different scenes. Existing methods, relying on fixed-dataset training, struggle to cope with continuously updating multi-domain vehicle data in real-world scenarios, and thus suffer significant performance degradation under data distribution shifts. While lifelong learning techniques can gradually accumulate multi-domain knowledge, they suffer from catastrophic forgetting of historical discriminative features and adapt inefficiently to new domains, particularly in preserving cross-domain spatial cues (e.g., component structure, contour details). To address this, this paper proposes a lifelong vehicle ReID model based on ResNet50, which embeds the depthwise spatial attention module. By leveraging depthwise separable convolution to dynamically focus on cross-domain invariant spatial regions such as the grille, headlights and vehicle contours, the model enhances local discriminative features to mitigate feature drift. The model couples two mechanisms: the depthwise spatial attention module preserves historical discriminative spatial features and curbs catastrophic forgetting, while the feature alignment loss maintains consistent multi-domain representations and balances old and new knowledge. Extensive experiments on VeRi-776, VRIC, and VehicleID show that, among the lifelong methods compared under a common protocol, the proposed method attains the highest accuracy—77.0% mAP and 78.7% R1 on average—while storing no historical images and requiring the lowest peak training memory of 7.2 GB.
// Source
Authors: Leilei Rong, Lisu Han, Lu Bai
Institutions: Weifang Medical University, Weifang University, Weifang University of Science and Technology