Searching for rare traffic signs

Abstract

Deep neural networks are good at recognizing traffic signs when they are trained on many different examples of each one. However, some traffic signs are very rare and not often encountered when collecting data. This means that a network does not recognize rare traffic signs as well as those that are encountered often. When collecting large amounts of data, one usually only labels a small subset of it. Therefore, there might exist more examples of the rare traffic signs in the unlabeled data set. If these examples could be found and used in training, the performance of the model could be expected to improve. This thesis evaluates how a standard neural network performs in searching for rare traffic signs, and whether some commonly used techniques from few-shot learning can improve its performance. To our surprise we find that they cannot. Furthermore, in this thesis we show that searching for rare traffic signs is an efficient active learning method, outperforming other established methods by requiring up to 8x less additional data to achieve the same F1-score on rare traffic sign recognition.