Synthetic Training Data Generation and Domain Randomization for Object Detection in the Formula Student Driverless Framework

Today industries strive toward using data-driven machine learning wherever applicable. Consequently, they re-quire manually or automatically labeled training data sets. Currently, synthetically generating labeled training data sets belongs to the open challenges in machine learning across multiple application fields. In this paper, we propose employing a procedural pipeline combining BlenderProc with domain randomization to create prelabeled training data sets synthetically. Randomizing the domain using uncorrelated random background images, we ensure that the neural network applied for object detection purely learns the object features and is background-independent. Our proposed pipeline yields a solution to create sizeable prelabeled training data sets. We assess the pipeline performance for the application of cone object detection for the formula student driverless competition using no real training and a small real-world training data set for fine-tuning: We show that using the synthetically generated training data fine-tuned with a limited real training data set performs best for object detection. This transfer learning-based, fine-tuned solution also outperforms the benchmark training data set in detecting knocked-ver cones that are neither present in the real nor the synthetic training data set. Consequently, by combining BlenderProc and domain randomization, we provide a solution for formula student teams to generate extensive training data for cone detection and other detection problems relevant to driverless.