The problem
A tier-one supplier needed driver state data against a homologation deadline. Their internal set had inconsistent event boundaries, so temporal models trained on it behaved unpredictably.
What we did
We defined onset and offset conventions for each state before any annotation started, then measured boundary agreement explicitly and reported it per state.
Infrared and RGB were annotated to the same schema so the fusion model did not need a mapping layer.
Outcome
The supplier hit the homologation date with a set whose boundary consistency they could document.
