Consistency-guided model-contrastive federated learning for scalable road damage detection

Machine learning has been increasingly adopted to automate road damage detection by analyzing widely available street-level images and videos using object detectors. However, most of these detectors are developed using centralized learning, where data from diverse sources must be consolidated and processed at a single location, creating high data transmission and storage costs and significant privacy concerns. This paper proposes a federated learning method that enables collaborative development of object detectors across different entities without requiring data sharing. The core idea is to learn a generalized feature extractor that captures discriminative, robust representations of road damage. This is realized by encouraging cross-domain consistency through swapping-based knowledge distillation and by aligning local learning objectives with collective knowledge through multi-scale model-contrastive learning. Extensive experiments on a multi-country dataset demonstrate the superior performance of the proposed method, especially in the face of data heterogeneity that restricts conventional federated learning methods in real-world applications.