1 | initial version |
we cannot help you with any unsupported 3rdparty c# wrappers, so your code is somewhat irrelevant here.
I have single image of most of the users
no, you can only expect something valid with 20+ images per user
it should not recognize the user for which it is not trained so far.
that's probably asking for too much. all it does is retrieve the closest item from the database, come rain or come shine. there is no real provision for "unknown" persons. all you can try is to check, if the distance is beyond a certain threshold.
besides that, you should probably use facenet (which is trained on triplet loss) instead of those classes, nowadays.
2 | No.2 Revision |
we cannot help you with any unsupported 3rdparty c# wrappers, so your code is somewhat irrelevant here.
I have single image of most of the users
no, you can only expect something valid with 20+ images per user
it should not recognize the user for which it is not trained so far.
that's probably asking for too much. all it does is retrieve the closest item from the database, come rain or come shine. there is no real provision for "unknown" persons. all you can try is to check, if the distance is beyond a certain threshold.
also, proper preprocessing (lighting, cropping, alignment, etc) is essential here.
besides that, you should probably use facenet (which is trained on triplet loss) instead of those classes, nowadays.
3 | No.3 Revision |
we cannot help you with any unsupported 3rdparty c# wrappers, so your code is somewhat irrelevant here.
I have single image of most of the users
no, you can only expect something valid with 20+ images per user
it should not recognize the user for which it is not trained so far.
that's probably asking for too much. all it does is retrieve the closest item from the database, come rain or come shine. there is no real provision for "unknown" persons. all you can try is to check, if the distance is beyond a certain threshold.threshold. (and again, you need more than 1 image for that)
also, proper preprocessing (lighting, cropping, alignment, etc) is essential here.
besides that, you should probably use facenet (which is trained on triplet loss) instead of those classes, nowadays.