In the Search for the Balance Between Real and Synthetic Images in Multi-Class Detection Systems

Intelligent systems focused on traffic management have been in evidence in recent years, and applications related to vehicle detection and tracking, speed estimation, and traffic flow identification have become an interesting research topic. For the previously mentioned tasks, a large number of data has to be gathered to train deep learning algorithms, but collecting that data can be a time and resource-consuming task. Therefore, the use of synthetic data has become a viable option that helps to minimize data acquisition problems, but when misused, it can negatively impact the model's quality. This paper presents a systematic literature review based on the use of synthetic images to train object detection models in urban scenarios, aiming at identifying the ideal ratio between real and synthetic images that can benefit those models and the best methods to produce synthetic images. This study identified that there is no consensus on the number of synthetic images that can help to generate a more accurate model, due to the low number of papers addressing this relationship, however, it was noted that the use of generative adversarial networks (GANs) can create synthetic images that are more similar to real images, bringing benefits for training detection models, although without identifying how the use of images generated by this method can help in the relationship between synthetic and real.