Face detection require a minimum number of visible attributes (2 eyes etc)

[bagobones]

I believe both FaceNet and ArcFace support face attributes (eyes, nose, mouth). I have a household of faces trained I am rarely getting false positives now other than obscured or faces off angle in profile.

Someone in almost full profile (one eye visible) or their face is coverage by a phone / hand / hair seems to cause a greater degree of false positives.

I wonder like the blur filter we could set a min threshold on visible attributes for labeling or reduce the confidence for obscured faces?

[NickM-27]

I believe both FaceNet and ArcFace support face attributes (eyes, nose, mouth)

Not directly. The face landmark detector could be used for this but in most cases it detects 2 eyes even if only one is visible just based on typical distances.

[bagobones]

Your right both eyes continue to show a dot when looking at demos like the ones here https://insightface.ai/arcface it does seem it keeps both eyes tracked even in profile, however of note someone in full profile they overlap / touch or have very little distance.

I wonder if a MIN distance between eyes could help or eliminate when overlapping / touching?

The system is working quite well but the accuracy is clearly BAD for off angle images to the point I have had strangers match a few times.

[NickM-27]

I think that is a collaborative example, arcface model itself only returns the face embedding. It does not return the points for the location of the eyes.

[bagobones]

Found this based on OpenCV https://mittalparveen652.medium.com/face-eye-and-mouth-detection-with-python-458f2a028674

Its just an idea, that might help lower the scores of side profiles, backs of heads, things that aren't actually faces.