How Angle Affects Embedding Accuracy
Face recognition networks are trained on predominantly front-facing images because most real-world applications require matching front-facing portraits. When you provide a photo at an angle, the 2D projection of your 3D face changes, features that are prominently visible from the front are partially obscured at an angle, and vice versa. This changes the feature pattern the network extracts.
The three degrees of angular freedom, yaw (left/right rotation), pitch (up/down tilt), and roll (sideways tilt), each affect the result differently. Yaw (looking left or right) has the largest effect, because it significantly changes which features are visible. Pitch and roll have smaller effects at moderate values.
Optimal Angle Range
The optimal range for face matching is ±15 degrees of yaw (within 15 degrees of straight-on), ±10 degrees of pitch, and ±10 degrees of roll. Within this range, the face alignment pipeline can successfully normalise the crop, and the embedding remains stable. Beyond ±25–30 degrees of yaw, accuracy begins to degrade noticeably.
A slight upward tilt (camera slightly above eye level, looking up) is the most common angle in controlled portrait photography and is well represented in training data. Photos taken from dramatically below (looking down the nose) or very high above (making the face appear foreshortened) should be avoided.