Automatic Generation of Images Using Unreal Engine for Supervised Learning

Caio B. Brasil, Ryan Oliveira, Ailton P Oliveira, Carnot Braun, Lucas Silva, Ilan S Correa, Aldebaro Klautau

DOI: 10.14209/sbrt.2023.1570920088

Evento: XLI Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT2023)

Keywords: Computer vision synthetic data virtual world supervised learning

Abstract

Many applications of machine learning (ML) require a large amount of labeled data to be used in practical deployments. Collecting data with labels can be a laborious and time-consuming task. A sensible alternative that has been widely adopted recently is to use synthetic data, generated by simulations, and automatically create labels within this process. This paper uses this approach to generate realistic datasets for training ML models to be used in computer vision. The proposed methodology is based on 3D virtual environments created by the Unreal Engine, and geometric relations to properly position the bounding boxes corresponding to each object of interest. To validate the methodology, a dataset of 3000 labeled images was generated in 2.5 minutes. Using the YOLOV7 deep neural network, nearly 100\% of accuracy was achieved using a test set disjoint from the training data.

Download