CuneiForm Method for Assuring the Safety of ML-Based Computer Vision Development Datasets

In this research, we introduce the CuneiForm Abstract Characterisation method, aimed at bridging the semantic gap between textual requirements and image datasets for computer vision systems. Focusing on the case study of an Eagle Drone, we demonstrate how our CuneiForm-based approach can systematically define and validate a dataset's compliance with safety requirements. The CuneiForm characterisation process involves constructing abstract images that encapsulate objects and behaviours outlined in requirements, such as adversarial drone movements and optical appearances. These abstract images are then instantiated with actual images in training sets, ensuring a traceable link to the written specifications. Key features of our approach include meaningful mapping of real-world Target Objects of Interest (TOIs) and their environmental behaviours into pictorial models, mitigating perception failures due to TOIs' pictorial behaviours.