Scalable Deep Learning Framework for Public Transit Sign Detection

Abstract. Detecting public transit signs in urban environments is a complex and challenging task due to the significant variability of these signs across different regions and cities. Unlike standardized traffic signs, transit signs often vary considerably, requiring adaptable and robust solutions for effective detection. In this work, we propose a novel Domain-Specific Agnostic Segmentation Method with Contrastive Learning, integrated into a scalable two-phase detection pipeline. This innovative approach enables our model to adapt to city-specific visual patterns without manual prompts, achieving accurate detection even for small and distant transit signs. As a significant contribution, we introduce the first dataset dedicated to public transit signs, consisting of 2,300 manually annotated images from multiple European cities. Our method significantly outperforms existing Swin Transformer-based detection models in real-world tests. These results establish a new benchmark for public transit sign detection, highlighting the effectiveness of our approach in providing a robust, scalable solution for intelligent transportation systems and reducing the need for retraining in diverse urban environments.