Computerized methods for creating tracks of locations across frames of a video corresponding to a facial feature of a human. A set of feature location hypotheses is generated, as applied to images derived from the sequence of frames representing images of the human. Each hypothesis is refined, and a first set of confidence measures is associated with each hypothesis. A second set of confidence measures is associated with interframe transition, and a cost function that is a combination of hypotheses and transition confidence measures is minimized. A set of tracks is generated, characterizing each of a plurality of facial features within each frame of the sequence of frames. Performance analysis data may further be derived in a performance driven animation production pipeline, based on the generated tracks.