Validating a Low-Cost Depth Sensor for Facial Anthropometry: Capabilities and Appropriate Use

BACKGROUND: Low-cost depth sensors are increasingly proposed for 3D facial capture, with high intraclass correlation coefficients (ICC) cited as evidence of readiness. Intraclass correlation coefficient reflects relative agreement and can stay high even when absolute error is large. METHODS: Twelve adults were measured by 2 observers at 16 facial distances. We recomputed device-versus-caliper ICC(3,k), Bland-Altman statistics, observer reliability and measurement error, scan-rescan precision, mean absolute error (MAE), and a tolerance classification, against a structured-light scanner and direct calipers. RESULTS: Device-versus-caliper ICC was high for both classes (0.88 and 0.93). Despite this, the depth-sensor MAE was 2.75 versus 1.64 mm (P=0.0002), and ICC did not predict absolute error (rho=-0.17, P=0.52). This error matched manual landmark-placement error (SEM 1.6-2.1 mm), so it is only partly device-driven; the sensor was distinguished instead by poorer scan-rescan repeatability (0.84 versus 0.36 mm). On a tolerance scheme, 14 of 16 depth-sensor distances exceeded 2 mm, falling to 7 of 16 for the device-specific component. CONCLUSIONS: At a fraction of the cost, the depth sensor reproduces relative ranking well but is less precise on the device-only surface metric. It suits morphometric surveys, ranking, and within-subject monitoring, not high-fidelity reconstruction. A high ICC alone does not establish clinical usability; absolute error must accompany it.