Annotation error analysis
Data quality
The examination of labeling mistakes to understand their patterns and causes.
Group errors by type, severity and context before deciding how to fix them. A timestamp offset calls for a different correction from an ambiguous label definition. Looking only at an average error rate can hide a concentrated problem in a task, camera view or supplier batch.
In practice
A review finds that release events are consistently marked several frames before the object leaves the gripper.