The 478-Point Face Mesh: What the Model Actually Sees
Every geometry value Selfbaus reports begins with a set of points laid over your face by a neural network. Understanding what those points are, and what they are not, is the foundation for reading any facial measurement honestly.
Where the mesh comes from
Selfbaus uses the MediaPipe Face Landmarker, run on-device. Google's task documentation says its face mesh model outputs an estimate of 478 three-dimensional landmarks, and that it can also output a facial transformation matrix, which we use for head pose. Blendshape (expression) output is switched off in our build. The 468-point surface mesh was described by Google researchers as an approximate 3D mesh inferred from single-camera input for AR applications (Kartynnik et al., 2019). A later paper describes an attention-based architecture that concentrates on regions such as the eyes and lips (Grishchenko et al., 2020). MediaPipe itself is a general framework for building perception pipelines (Lugaresi et al., 2019). We have not verified which paper describes the exact network inside the bundle we ship, so we cite these as background, not as a specification of our model.
What a landmark is
A landmark is a prediction, and the original mesh was designed for AR effects, not for anthropometry. Kartynnik et al. state the 468 points were manually selected for applications such as AR effects, try-on and makeup, with denser points where human perception is more sensitive. The network was trained on images with annotated geometry, and for a new photo it estimates where each of its fixed, numbered points should sit. Point 152 always means "the chin-bottom point on the mesh". It does not mean the model has found your menton the way a clinician finds it by palpating bone.
That distinction matters most at the edges of the face. Points on the oval outline, including those near the forehead top and jaw sides, are positioned by the model from learned shape, not from a visible boundary. In particular, the top-of-oval point is not the hairline, and the width points at the sides are contour estimates, not the bony cheekbone width used in anthropometry.
What the mesh gives us, and what it does not
It gives dense, repeatable coordinates in the image plane, plus an estimated head pose. It does not give millimetres: a single photo has no scale. It does not give depth you can trust for profile measurements: the mesh has a z-coordinate, but it is the model's guess. So Selfbaus stays with frontal, 2D, ratio-style observations and says so.
Why we pin the model
Landmark positions change between model versions. Selfbaus 0.2.0 loads the official MediaPipe float16 v1 face_landmarker.task bundle (3,758,596 bytes, SHA-256 64184e229b263107bc2b804c6625db1341ff2bb731874b0bcc2fe6544e0bc9ff) through MediaPipe Tasks Vision 0.10.32 in the browser, with detection and presence thresholds of 0.8. Each report carries the engine and model identity, so a result can be reproduced and a change in numbers can be traced to a cause.
References
- Google AI Edge. Face Landmarker guide. https://ai.google.dev/edge/mediapipe/solutions/vision/face_landmarker
- Kartynnik, Y., et al. (2019). Real-time facial surface geometry from monocular video on mobile GPUs. https://arxiv.org/abs/1907.06724
- Grishchenko, I., et al. (2020). Attention Mesh: high-fidelity face mesh prediction in real-time. https://arxiv.org/abs/2006.10962
- Lugaresi, C., et al. (2019). MediaPipe: a framework for building perception pipelines. https://arxiv.org/abs/1906.08172