Master thesis : Is the use of synthetic datasets a solution to improve object detection models on real data?

Abstract

In the recent years, object detection models have leveraged deep learning architectures to improve performance in many problems. However, these techniques require a large amount of high quality labelled data in order to reach their full potential, and obtaining such data may prove to be an arduous task. In this context, this work explores the possibility of using entirely synthetically generated and labelled images to train an object detection model. In particular, we examine which factors of variations in the synthetic data best transfer to real data. Unsurprisingly, models trained on synthetic data only perform significantly worse than models trained on real data. We explore whether the synthetic images can be enhanced using filtering and generative models, but find the results to be inconclusive. In a setting where both real and synthetic data are available, we experiment to find out how these should be combined to improve performance in the real domain. We find that the synthetic and real datasets should be combined into a single training dataset, and that the object detection model trained in this fashion significantly outperforms the model trained on real data only.