Hungarian Traffic Sign Detection and Classification using Semi-Supervised Learning
Levente Kovács, Gábor Kertész · 2021
Semi-supervised learning is a special way to improve the classification performance of a model where labeled data are not available. By using unlabeled observations and handling them as training data in a transfer learning buildup, we get a structure often referred to as self-supervision. In case of traffic sign detection and classification the task is complicated to the large number of tables and the different representations from country to country. While a number of public datasets are available, these might not give satisfying performance; to deal with this issue, a semi-supervised method is presented where frames of dashcam recordings are automatically annotated and reused as training samples.