Head Pose: Why Yaw, Pitch and Roll Decide Whether a Photo Can Be Measured

A face is a three-dimensional object and a photo flattens it. Turn your head a few degrees and the flat image changes: one eye looks smaller, the nose seems to shift, the jaw looks narrower on one side. If you measure that image as though the head were square to the camera, the error ends up inside every ratio.

The three angles

• Yaw is turning left or right, like saying no.
• Pitch is nodding up or down, like saying yes.
• Roll is tilting the ear toward the shoulder.

Selfbaus reads these from the facial transformation matrix the MediaPipe Face Landmarker returns, which maps a canonical face model onto the detected face. The pose is a model output, with its own error.

Why a gate

Selfbaus currently requires yaw and pitch within 10 degrees and roll within 5 degrees. These limits are proposed. They are our judgment of "close enough to frontal" and we have not validated them against ground-truth pose data. The pose estimator's error on selfie-style photos is also uncalibrated, and we say that on the methodology page.

Roll is the easy one, and still not free

Roll rotates the image plane, so it can in principle be corrected by rotating the landmarks back. The symmetry measure does this by rotating to the eye line first (see the symmetry article). But a few degrees of estimation error is the same size as the effects many people care about, such as canthal tilt. Any small-angle measurement is therefore noisy at the single-photo level.

What to do about it

Hold the camera at eye level, look straight into the lens, and keep your chin level. If you want a number that means something, take several shots and look at the spread.

Limits

Pose gating reduces one error source. It does not remove lens, distance or expression effects, covered in the other articles in this section.