-
-
Notifications
You must be signed in to change notification settings - Fork 46
Mininimum face size
The face detector can discover very small faces, however these may be unclear for proper comparison. A general rule in neural networks, if the human can't recognize it, don't expect better results. If the face is very small, a human take additional information (hair, his clothes, who is it with, where was the photograph taken, etc.), and can easily infer who the person is.
Note that I said the hair, because this is not part of the face, and our detector does not take it into account. Let's see an example:
If you ever saw The Big Bang Theory you will easily recognize that this character is Sheldon Cooper. The rectangle shows the face detected by our default detector. As you can see, the photo has good quality, and the face a good size, 283x283. But what happens if we cut the face?
Surely you will still recognize the character, but observe how as you shrink your face, more information is lost. Many details in the eyes, the nose, are lost.
At first glance, up to 150x150px (73.77% in my tests with real photos) most details of the original photo are still visible. On the other hand, the neural network that we use for facial recognition, was trained entirely with faces of this size. Therefore we assume that it is the minimum size to obtain very good results. Using this value, in my case 26% of my faces will not compare (Still can see and rename these, but they are not clusteres by default), but I but this way I make sure that the clustered faces will have great quality.
The faces less than 125x125px (13.36% in my photos.) probably has enough detail to speak of a quality comparison. Up to 60px (2.32% in my tests), you can still recognize it, and our recognition model will still work well, but you will notice the lack of detail. If you have similar people, there are possibilities that start to get confused.
Going to the extreme, nobody can believe that with a 30x30px face you can reliably recognize someone. And these are the cases that we want to avoid with this setting.