面部检测检查面部用于量化面部质量问题

我正在考虑的一些快速问题，如果其中一些问题先前在其他地方得到解答，我道歉。 

Some quick questions I'm contemplating, if some of them are answered previously elsewhere, I apologize. 

我正试图量化一张脸的质量（得分是一个很好的图像框架使用？）。  模糊，不可思议，倾斜，角度等，从而"得分"。面子的类型。 

I'm trying to quantify the quality of a face (score is this a good image frame to use?).  Blurry, uncanny, tilt, angles, etc. and thus "score" the type of face. 

首先，我正在考虑使用 

First, I'm considering using 

faceRectangle  - 获取尺寸并忽略较小的面孔。  

faceRectangle - get the size and disregard smaller faces.  

第二名： 

2nd: 

analyzeFaceLandmarks

analyzesFaceLandmarks

如果找不到一个或多个地标会发生什么情况，那么面部检测会一起失败？ 什么是用于delta的好地标，例如 

What happens if one or more of the landmarks is not found, then the face detect fails all together?  What are good landmarks to utilize for delta like 

if ( f.eyeRightInner.X - f.eyeLeftInner.y <20) == bad (not enough distance between inner eye)
if ( f.eyeRightInner.y - f.eyeLeftInner.y >10 ) == bad (headPose.pitch same thing?) 
if ( f.upperLipBottom.y - f.underLipTop.y >5)  == mouth is open ? 

face.attributes.headPose:

"headPose": { "pitch": 0.0, "roll": -10.3, "yaw": 18.1 },

为了获得牛津ML AI的最佳质量培训，我们是否应该提供32个不同角度的面孔，
姿势等？ 

将一个人的faceIds培训成人员组的最佳面部结果或类型是什么？ 

faceId是GUID吗？ 

What are the best face results or types for training a person's faceIds into a PersonGroup set? 
Is faceId a GUID? 

"faceId": "c5c24a82-6845-4031-9d5d-978df9175426"

谢谢。

肯