scoring

[pgg]

for FaceResult what is the difference between:

score, boxScore and faceScore? more specifically, when detecting a face which is used to determine the quality of the detection? Especially if used for later similarity calculations. The use case here is enrolling a face via detect, make sure it is a fairly good match (score or boxScore or faceScore above required threshold?). Then later using the enrolled embedding to determine similarity . . .

[vladmandic]

boxScore is score of face detection in overall input.
faceScore is score of face keypoint detection in cropped face and can be missing if facemesh model is disabled.
score is simply faceScore if its available and if not, then boxScore.

now, you could enroll face in similarity calc without ever running through facemesh,
but running through facemesh does have two advantages:

• it performs slightly more precise face cropping
• it can optionally perform rotation corrections (if rotation is enabled in config)

neither are not required, but they do help to determine better face descriptors which are used for similarity checks.

[pgg]

ok, thanks!

[plmok61]

Hi @vladmandic,
I am trying to better understand faceScore and how it should be used. Can this score seen as a confidence that there is a real face in the image? For example if we have above 0.5 for faceScore, we can be pretty confident that we are seeing a real face or am I misunderstanding this value?

[vladmandic]

faceScore is confidence of detection of facial keypoints, nothing to do with real or not.