Face_recognition: How to scale face_recognition?

[sskgit]

• face_recognition version: latest
• Python version: 2.7
• Operating System: MacOS

This is a great repo!

Installed face_recognition and tried all examples (except webcam). So far works great.

While trying to do this on 10 unknown images and one known images:

face_recognition ./known_people/ ./unknown_pictures/

it takes around 15-20 seconds. Is there a way to do image comparison of one unknown image to large number of known images in ~2-3 seconds. Also, will this code run on GPU instance and will it reduce latency?

[ageitgey]

If you compile dlib manually with CUDA support (after having installed CUDA and cuDNN), it will use the GPU for generating the face embeddings. So that's a bit faster. However the part that still isn't GPU-accelerated is finding the locations of the faces in each image.

Also the face_recognition script isn't caching anything, so it's rescanning all the known_images each time you run the program which is slower than it needs to be.

One way to speed things up a little bit would be to change the script to pre-cache the face embeddings for each known image and save them somewhere:

1. Cache the results of running face_recognition.face_encodings(unknown_image)[0] on each known image. You only need to do that once.
2. When you want to test an unknown picture, load those cached results into an array called cached_known_faces or something and pass that to face_recognition.compare_faces([cached_known_faces], unknown_face_encoding).

I want to work on improving performance in general, but I haven't had a chance yet.

[sskgit]

Thanks @ageitgey for the clarification.