I have found something odd.
I have a database of categorized face images. I can register these faces, as I have deleted all the ones that come back with “no face detected” when registered.
If I then identify the face in each of these images, and extract the face into a new database (with an added margin of 30%) using the returned co-ordinates. A large number of these faces will not register - they come back as “no face detected”.
So it seems that if you extract the faces from a database using Deepstack (even with a margin), then try to register just the face (without all the background), then a large number of faces will fail to register.
Why is this? why will Deepstack not register a face it has just identified as a face?
Here is an example
This face registers:
This face does not:
The face that does not register is extracted from the first face, with a 30% margin all round. The only difference I can see is that the extracted face is 177X200 .png and the original is 200x200 .jpg, so it’s slightly cropped in the horizontal direction.