Methodological Guidelines for Model ODD Characterization

This guide presents a rule-based and top-down approach to characterize the ODD from a prediction model and its training dataset. This approach can be cut in several steps: choice of anomalies (ODD parameters), definition of anomalies levels, creation of datasets containing those anomalies levels, inference on those datasets and inference results analysis from predefined rules to characterize zones of anomalies levels where the model prediction can be trusted or not.