Temporal consistency
Physical AI
Consistency in identities, labels and events across a sequence over time.
For video and robot data, review can check that tracks, action labels and time boundaries agree with the observed sequence. A series of plausible individual frames can still form a contradictory annotation. Define tolerances for timing and specify how gaps or occlusions should be handled.
In practice
An object should not switch identity between adjacent frames without evidence of a real change.