Evaluating a Siamese Network for Traffic Sign Recognition Using Synthetic Datasets

Traffic sign recognition (TSR) is essential for au-tonomous vehicles and advanced driver assistance systems (ADAS). Machine learning methods often require large labeled datasets, limiting their use when available data is scarce. One-shot learning addresses this issue by enabling systems to learn new classification categories with a minimal set of examples. This paper adapts a Siamese neural network architecture for the task of traffic sign recognition, capable of learning from only tens of samples per class. We propose and utilize image augmentation methods to create synthetic datasets based on Slovak road sign templates. Finally, we evaluate and compare the classification performance of the Siamese network model on synthetic and real-world datasets from different countries.