Confidence
The confidence within this application refers specifically to the confidence of the detection of a face during the analysis of a particular image. Blurred or misaligned faces would have a confidence closer to 0.0, and the best images close to 1.0. We take this value to evaluate it as a parameter that suggests the quality of the face and its corresponding Descriptor.
The detection of a face and the obtaining of its Descriptor are two independent processes. Confidence refers to the first process. It's just a score that says how reliable the detection is. On the other hand, obtaining the Descriptor of the face always returns an array of 128 numbers, which depend mainly on the training of the model, and the image used, but the quality of this descriptor cannot be evaluated. In short, you could pass anything to the model and always return an valid array of 128 numbers.
How we use this Descriptor to group similar faces, we must set some parameter to know if this comparison can be reliable. We can assume that if a face is blurred -Which results in a low confidence value- the Descriptor will not be so reliable, and we should not use it. This is similar to what any real person does. Looking at a blurred photo, can you say 100% sure who is the person in it?. On the other hand, if the photo is good, the confidence will be high, and we can assume that the Descriptor is reliable and we can compare it to obtain groups of similar faces.
- This setting determines how reliable must be a face detection to try to group it.
Any face with a confidence less than this value is not compared to other faces, and therefore it will be stored as a single group.
Using a high value close to 1 only very reliable faces will be grouped, resulting in many single groups, but also ensuring that the largest groups are all of good quality, and probably more reliable.
Using a low value close to 0, the process will attempt to group all faces, resulting in few singles groups but also the larger groups will have faces of poor quality mixed, and there are probably several persons mixed.
