Traffic sign repositories: bridging the gap between real and synthetic data

Abstract

Creating a traffic sign dataset with real data can be a daunting task. We discuss the issues and challenges of real traffic sign datasets, and evaluate these issues from the perspective of creating a synthetic traffic sign dataset. A proposal is presented, and thoroughly tested, for a pipeline to generate synthetic samples for traffic sign repositories. This pipeline introduces Perlin noise and explores a new type of noise: Confetti noise. Our pipeline is capable of producing synthetic data which can be used to train models producing state of the art results in three public datasets, clearly surpassing all previous results with synthetic data. When merged or ensemble with real data our results surpass previous state of the art reports in three datasets: GTSRB, BTSC, and rMASTIF. Furthermore, we show that while models trained with real data datasets perform better in the respective dataset, the same is not true in general when considering other similar test sets, where models trained with our synthetic datasets surpassed models trained with real data. These results hint that synthetic datasets may provide better generalization than real data, when the testing data is outside of the distribution of the real data.