Generating Synthetic Data Using a Knowledge-based Framework for Autonomous Productions

摘要

Instead of taking images with a camera, synthetic data is generated in a computer simulation. One advantage of this is that training data can be generated on-demand, e.g., to automatically retrain robots when a task changes. While this makes synthetic data a promising approach for autonomous productions, realizing such autonomous setups is difficult with current systems for generating synthetic data, which usually require a programmer for every dataset to be generated.To overcome this problem, we present a novel framework for generating synthetic data. This framework restructures the generation process into asynchronous phases to increase the level of autonomy in two ways. First, by letting programmers write parameterized scripts, many different datasets can be autonomously generated. Secondly, by introducing a user interface, domain experts are enabled to influence the generation process on their own without a programmer. Furthermore, by being built as a new layer on top of existing systems for generating synthetic data, our framework shows a new way to maximize compatibility with other research on synthetic data generation.To test our framework, we have developed a fully functional prototype based on it. Successfully using this prototype for an example experiment, we conclude that our ideas work. Future research can use our prototype for more elaborate experiments on autonomous productions and to further assess its usability.