Fuzzy Database for Language-Driven Procedurally Generated Simulated Datasets

Current generation artificial intelligence (AI) and machine learning (ML) heavily rely on data. One way to address this bottleneck is through simulation. In previous work, we presented a scalable procedure to automatically generate a large volume of diverse simulated scenes and images, along with their corresponding ground truth, based on a formal language description. However, this approach assumes the ability to establish connections between language and the underlying simulation assets, such as materials (e.g., textures) and geometric models (e.g., 3D models). In this article, we discuss a process and user interface for annotating assets to create a fuzzy database while considering associated uncertainties. Once the database is constructed, user language statements can be employed to query our assets. Each query involves aggregating relevant criteria, resulting in a membership degree and a ranked order of assets. Ultimately, these values are used by our procedural algorithm to sample new scenes. In summary, the proposed methodology offers an efficient approach to catalog, manage, and query assets for AI/ML training and evaluation in simulation.