Synthetic Data for Object Detection with Neural Networks: State of the Art Survey of Domain Randomisation Techniques

Abstract

Machine learning relies heavily on access to large and well-maintained datasets. In this article, we focus on Computer Vision and object detection applications to survey past research on automatic generation of annotated datasets that does not require costly and time-consuming human labelling. In specific, we analyse research done in the area of Domain Randomisation applied to Neural Networks predominant in object detection since the last decade. We propose a set of criteria for comparison of previously published works, and utilise these criteria to make conclusions about various trends in the area, similarities/differences and key discoveries made since conception. The purpose of this work is to advise practitioner on leading solutions and help researchers gain better understanding of the landscape. The key takeaways from our analysis show the current state of the art solutions within the mid-quartile range allow object detection with typically about 1-25% performance decrease in comparison to manually annotated datasets; while the top performant approaches above the upper quartile gain about 2-32% lead over real data training in their specific application areas. Our survey shows the future outlook is more research into 3D generation techniques, with most innovative yet complex techniques related to end-to-end modifications of entire network architectures to suit synthetic data training.