Minimizing Gaps Between Synthetic And Real Datasets

Abstract

Creating synthetic images that are of high quality is a crucial step for many deep learning projects, especially when real data are either limited or too expensive to acquire. Despite its importance, there has been a noticeable gap in the exploration of efficient techniques in the synthetic image generation field, except for generating photo-realistic samples of high cost. In this study, we bridge this gap by investigating the key elements and features that should be maintained and aligned in synthetic datasets to mirror the properties of real datasets. From our findings, we have developed practical guidelines that simplify the process of creating synthetic datasets that can stand up to their real counterparts, particularly in the realm of object detection tasks. We create a reusable framework that can guide future research and developments in this area.