Corner case
How it works
In engineering practice, corner cases are identified and handled via: (1) systematization and taxonomy of types (e.g. pixel-, domain-, object-, scene- and scenario-level anomalies for visual perception); (2) deliberate collection and labeling of hard scenarios (dedicated datasets, e.g. road corner cases); (3) corner-case / out-of-distribution and anomaly detection at runtime; (4) scenario and simulation testing covering simultaneous extreme conditions. The goal is to shrink the long tail of failures by covering it with data and safety mechanisms.
Problem solved
Names and organizes the problem of rare, unforeseen out-of-distribution scenarios that make AI systems fail despite strong average-case performance โ providing a framework for their systematization, detection, and validation.
Implementation
Corner cases are rare by definition and barely appear in collected datasets, hampering training and evaluation.
Tuning only for known corner cases fails to generalize to unseen long-tail cases.
Evolution
Breitenstein et al. present detection approaches for corner cases in visual perception for automated driving.
Bogdoll et al. define goals and challenges of describing corner cases, consolidating terminology.
CODA is released, a real-world road corner case dataset for object detection in autonomous driving (ECCV 2022).