Facial landmark detection accuracy of MediaPipe FaceMesh on Southeast Asian pediatric faces: Age-stratified NME analysis, ethnic-diversity implications, and protocol recommendations for clinical morphometry
DOI:
https://doi.org/10.54939/1859-1043.j.mst.114.2026.167-174Keywords:
Facial landmark detection; MediaPipe FaceMesh; Normalized mean error; Pediatric morphometry; Southeast Asian; Age-stratified benchmark; Ethnic diversity; Autism spectrum disorder; Clinical morphometry.Abstract
MediaPipe FaceMesh, the most widely deployed open-source facial landmark detector, was trained on adult faces, leaving its accuracy on pediatric Southeast Asian populations uncharacterized. This paper presents the first age-stratified Normalized Mean Error (NME) benchmark on Vietnamese children aged 2–8 years (n = 100; 14 landmarks; ICC ≥ 0.90). NME decreased significantly with age — 7.3%, 4.8%, and 3.1% IOD at 2–3, 3–5, and 5–8 years, respectively (ANOVA: F = 18.4, p < 0.001) — versus the adult benchmark of 2–4% IOD. Iris centers (lm468/473) achieved 1.0–1.5% IOD at all ages and are recommended as the preferred IOD normalizer. Philtrum-base (lm0) NME of 12.3% at 2–3 years renders philtrum height unreliable in toddlers; substituting outer lip corners (lm78/308) reduces philtrum width uncertainty by 40%. A first-order error-propagation analysis translates these per-landmark errors into biomarker-level reliability tiers benchmarked against published ASD effect sizes, providing a reusable framework for assessing morphometric biomarker reliability. Vietnam's multi-ancestry ethnic diversity introduces a further uncharacterized source of systematic error: admixture evidence predicts craniofacial geometries in Khmer, Cham, and highland Austroasiatic children lying partially outside FaceMesh's training distribution. Evidence-based landmark selection protocol recommendations are provided for pediatric Southeast Asian morphometric studies.
References
[1]. K. Aldridge et al., “Facial phenotypes in subgroups of prepubertal boys with autism spectrum disorders are correlated with clinical phenotypes”, Mol. Autism, Vol. 2, p. 15, (2011). DOI: 10.1186/2040-2392-2-15.
[2]. M. Boutrus et al., “Increased facial asymmetry in autism spectrum conditions is associated with symptom presentation”, Autism Res., Vol. 12, No. 11, pp. 1774–1783, (2019). DOI: 10.1002/aur.2193.
[3]. H. M. Özgen et al., “Minor physical anomalies in autism: A meta-analysis”, Mol. Psychiatry, Vol. 15, No. 3, pp. 300–307, (2010). DOI: 10.1038/mp.2008.75.
[4]. B. Hallgrímsson et al., “Automated syndrome diagnosis by three-dimensional facial imaging”, Genet. Med., Vol. 22, pp. 1682–1693, (2020). DOI: 10.1038/s41436-020-0845-y.
[5]. T. V. Thanh et al., “Detecting autism using facial images of Vietnamese children with deep learning techniques”, Int. J. Electr. Comput. Eng. (IJECE), Vol. 15, No. 5, pp. 4762–4773, (2025). DOI: 10.11591/ijece.v15i5.pp4762-4773.
[6]. H. A. Tung et al., “Impact of training data composition on Vietnamese autism facial screening: A factorial analysis of pretraining strategy and local data augmentation”, IAES Int. J. Artif. Intell. (IJ-AI), Vol. 15, No. 1, pp. 766–779, (2026). DOI: 10.11591/ijai.v15.i1.pp766-779.
[7]. V. Kartynnik, A. Ablavatski, I. Grishchenko and M. Grundmann, “Real-time facial surface geometry from monocular video on mobile GPUs”, in Proc. CVPR Workshops, Long Beach, CA, (2019).
[8]. C. Sagonas, G. Tzimiropoulos, S. Zafeiriou and M. Pantic, “300 faces in-the-wild challenge: The first facial landmark localization challenge”, in Proc. IEEE ICCV Workshops, Sydney, pp. 397–403, (2013).
[9]. Y. Wu and Q. Ji, “Facial landmark detection: A literature survey”, Int. J. Comput. Vis., Vol. 127, No. 2, pp. 115–142, (2019). DOI: 10.1007/s11263-018-1097-z.
[10]. I. Grishchenko, A. Ablavatski, Y. Kartynnik, K. Raveendran and M. Grundmann, “Attention mesh: High-fidelity face mesh prediction in real-time”, arXiv:2101.01691, (2021).
[11]. T. K. Koo and M. Y. Mae, “A guideline of selecting and reporting intraclass correlation coefficients for reliability research”, J. Chiropr. Med., Vol. 15, No. 2, pp. 155–163, (2016). DOI: 10.1016/j.jcm.2016.02.012.
[12]. A. Lu and M. Perkowski, “Deep learning approach for screening autism spectrum disorder in children with facial images and analysis of ethnoracial factors”, Brain Sci., Vol. 11, No. 11, p. 1446, (2021). DOI: 10.3390/brainsci11111446.
[13]. Y. Gurovich et al., “Identifying facial phenotypes of genetic disorders using deep learning”, Nat. Med., Vol. 25, pp. 60–64, (2019). DOI: 10.1038/s41591-018-0279-0.
[14]. L. Echeverry-Quiceno et al., “Population-specific facial traits and diagnosis accuracy of genetic and rare diseases in an admixed Colombian population”, Sci. Rep., Vol. 13, (2022). DOI: 10.1038/s41598-023-33374-x.
[15]. J. Chacón-Duque et al., “Latin Americans show wide-spread Converso ancestry and imprint of local Native ancestry on physical appearance”, Nat. Commun., Vol. 9, (2018). DOI: 10.1038/s41467-018-07748-z.
[16]. P. Claes et al., “Genome-wide mapping of global-to-local genetic effects on human facial shape”, Nat. Genet., Vol. 50, pp. 414–423, (2018). DOI: 10.1038/s41588-018-0057-4.
[17]. M. Lipson et al., “Ancient genomes document multiple waves of migration in Southeast Asian prehistory”, Science, Vol. 361, pp. 92–95, (2018). DOI: 10.1126/science.aat3188.
[18]. P. Changmai et al., “Indian genetic heritage in Southeast Asian populations”, PLoS Genet., Vol. 18, No. 3, (2021). DOI: 10.1101/2021.01.21.427591.
[19]. D. Liu et al., “Extensive ethnolinguistic diversity in Vietnam reflects multiple sources of genetic diversity”, Mol. Biol. Evol., Vol. 37, No. 9, (2020). DOI: 10.1093/molbev/msaa099.
[20]. E. Macholdt et al., “The paternal and maternal genetic history of Vietnamese populations”, Eur. J. Hum. Genet., Vol. 28, pp. 636–645, (2019). DOI: 10.1038/s41431-019-0557-4.
[21]. W. Kutanan et al., “Reconstructing the human genetic history of Mainland Southeast Asia: Insights from genome-wide data from Thailand and Laos”, Mol. Biol. Evol., Vol. 38, No. 8, pp. 3459–3477, (2020). DOI: 10.1093/molbev/msab124.
[22]. L. Le et al., “Norms of soft tissue cephalometric in the 18–25 age demography in Vietnam”, J. Orthodontic Sci., Vol. 12, p. 37, (2023). DOI: 10.4103/jos.jos_105_22.
[23]. Judith E. Allanson et al., “Elements of Morphology: Standard Terminology for the Head and Face”, American Journal of Medical Genetics Part A, Vol. 149A, No. 1, pp. 6–28, (2009). DOI: 10.1002/ajmg.a.32612.
[24]. Thuy Trang Thi Ho et al., “Dental-craniofacial Characteristics of Southern Vietnamese People with Well-balanced Face on Cephalometric Films and Its Comparison with Caucasians and Northern Vietnamese Population”, J. Int. Soc. Prev. Community Dent., Vol. 11, No. 3, pp. 316–323, (2021). DOI: 10.4103/jispcd.JISPCD_13_21.
[25]. S. Yucer et al., “Racial bias within face recognition: A survey”, ACM Comput. Surv., Vol. 55, No. 13s, (2022). DOI: 10.1145/3705295.
[26]. J. Buolamwini and T. Gebru, “Gender shades: Intersectional accuracy disparities in commercial gender classification”.
[27]. J. Huang, G. Galal, M. Etemadi and M. Vaidyanathan, “Evaluation and mitigation of racial bias in clinical machine learning models: Scoping review”, JMIR Med. Inform., Vol. 10, No. 5, (2022). DOI: 10.2196/36388.
[28]. R. Hartley and A. Zisserman, “Multiple View Geometry in Computer Vision”, 2nd ed., Cambridge University Press, Cambridge, UK, (2004).
[29]. O. Pavlovič et al., “Congruence in European and Asian perception of Vietnamese facial attractiveness, averageness, symmetry and sexual dimorphism”, Sci. Rep., Vol. 13, p. 14090, (2023). DOI: 10.1038/s41598-023-40458-1.
[30]. D. Lee et al., “High heritability of human facial traits reveals associations with CNTLN, BRCA1, and TMPRSS6 loci in Korean families”, Heliyon, Vol. 10, No. 1, p. e39173, (2024). DOI: 10.1016/j.heliyon.2024.e39173.
[31]. H. V. Huy, N. T. T. Tan and N. Q. Tao, “Vietnamese text recognition in scene images using deep learning”, J. Mil. Sci. Technol., Vol. 90, pp. 140–149, (2023). DOI: 10.54939/1859-1043.j.mst.90.2023.140-149.
[32]. N. V. Tra, V. C. Thanh and D. V. Sang, “Improving radar target recognition based on generative adversarial network”, J. Mil. Sci. Technol., Vol. 93, pp. 12–18, (2024). DOI: 10.54939/1859-1043.j.mst.93.2024.12-18.
