Build Your Own Training Data -- Synthetic Data for Object Detection in Aerial Images

Abstract

Machine learning has become one of the most widely used techniques in artificial intelligence, especially for image processing. One of the biggest challenges in developing an accurate image processing model is to collect large amounts of data that are sufficiently close to the real-world scenario. Ideally, real-world data is therefore used for model training. Unfortunately, real-world data is often insufficiently available and expensive to generate. Therefore, models are trained using synthetic data. However, there is no standardized method of how training data is generated and which properties determine the data quality. In this paper, we present first steps towards the generation of large amounts of data for human detection based on aerial images. To create labeled aerial images, we are using Unreal Engine and AirSim. We report on first impressions of the generated labeled aerial images and identify future challenges-current simulation tools can be used to create realistic and diverse images including labeling, but native support would be beneficial to ease their usage.