Object Detection Using Synthetic Training Data
Assar Pettersson, Filip Stjernström · 2020
Annotating data for machine learning purposes can be very inefficient and time-consuming. In this paper we introduce a pipeline for generating a 3d scene from a simple image. We discuss and develop the first step which includes a 3d network being able to recognise any particular object. We focus on how synthetic data can be used to make the annotating process simpler. We use generated synthetic images and train two different networks (YOLO and DOPE) and study their performances in order to learn how to create better training data. Finally, we conclude that domain randomization is very useful for attaining good results and discuss the diculties when studying training data.