On synthetic and real images as training data for object detection -- A brief review

Abstract

To train neural networks, sufficiently large and diverse datasets are needed. To address this, the use of synthetic data has become popular because it is inherently scalable and can be automatically annotated. A brief overview of recent work on using synthetic and real images as training data for object detection is presented in this paper, with a focus on mixing real and synthetic training data. The trend is that having real data and adding some amount of synthetic data helps the performance in many studies. It was concluded that there appears to be no consensus on the ratio of real and synthetic image data.