Object substitution-based contextual domain randomization for generating neural network training data

Domain randomization is a data synthesis approach for training neural network models on synthetic images for the purpose of generalization to real-world images through randomizing parameters of the rendering process such as lighting, backgrounds, object poses, and object textures, leading to improved performance in image-based neural network applications. Domain randomization increases the amount of variability in synthetic data to match the amount of variability that exists in real-world data. In this paper, we present a novel substitution-based domain randomization approach, object substitution-based contextual domain randomization (OSCDR), that preserves the contextual and structural integrity of real-world data. This approach is expected to enable the synthesis of cheap, high quality, and highly variable datasets for image-based neural network applications, including but not limited to object detection and localization, autonomous driving, and robot navigation.