3D facial landmarks: Inter-operator variability of manual annotation

Area of Science:

• Medical Imaging
• Biometrics
• Computer Vision

Background:

• Manual landmark annotation in medical imaging introduces significant variance, impacting results.
• Variability of human facial landmarks is under-researched compared to other fields like orthodontics.
• Current methods lack standardized procedures for facial landmark annotation, leading to inconsistencies.

Purpose of the Study:

• To present a comprehensive 3D facial annotation procedure.
• To identify a sparse set of manually annotated landmarks to reduce operator time and minimize variance.
• To establish a reliable method for dense point correspondence mapping in 3D facial scans.

Main Methods:

• Utilized 3D facial scans from 36 participants.
• Six operators performed dual manual annotations of 73 landmarks using a three-step scheme.
• Analyzed intra- and inter-operator variability using mixed-model ANOVA and compared four sparse landmark sets.

Main Results:

• Eye landmarks, especially pupil centers, showed the lowest variance; jaw and eyebrow landmarks exhibited the highest.
• Intra-operator variability and portrait variations had minimal impact.
• Using a sparse set of 14 landmarks reduced dense point mean variance from 1.92 mm to 0.54 mm.

Conclusions:

• Inter-operator variability is linked to specific, less defined landmarks.
• Operator training and portrait characteristics had minor effects on landmark variability.
• A sparse set of 14 landmarks effectively reduced variability and enabled creation of a dense correspondence mesh for facial feature capture.