Towards Knowledge-based Generation of Synthetic Data by Taxonomizing Expert Knowledge in Production

Abstract

Synthetic data is a promising approach for industrial computer vision because it can enable highly autonomous production processes. However, this potential is not fulfilled by current software for synthetic data generation, which usually requires a programmer to create new datasets. To overcome this, we are proposing a framework for more autonomous synthetic data generation, formalizing user roles relevant to such systems. A central aspect of our framework is that domain experts can easily influence the generation of synthetic data by entering knowledge via user interfaces. To get a better idea of what such knowledge could be, we have systematically collected examples of knowledge types for synthetic data generation in production and combined them into a taxonomy with almost 300 nodes. Using this taxonomy as the basis for analyses, we derive six implications for our framework, such as knowledge being not only passed on by domain experts but also by the designer of the user interfaces and generation algorithms. We plan to incorporate these findings to further refine and implement our framework in future research.