Face Identification on Jetson Nano

The python face_recognition library is a package that can look at a frame, find the faces, encode the found faces, and then compare found face encodings with an array of known face encodings. With a few lines of code it can not only find faces, but accurately determine the identity of the person. The issue is that it appears to not be GPU-friendly, and looks like it runs on the CPU.

Is there a GPU/Cuda friendly equivalent of this package? I know that with the jetson utility and inference packages you can load a frame, and find faces (facenet) quickly, and then display boxed face, but is there a tool that will do the face encoding and comparison on the GPU. If not, what would the path forward be . . . would you do transfer learning, and create a class for each person, and then train on the known faces?

Having a lot of fun learning the Jetson Nano, but want to figure out how to begin to better leverage the GPU. Appreciate any help on this.
Paul McWhorter

Hi,

The face_recognition python module use dlib as backend library.
And yes, dlib support GPU mode and Jetson platform.

Here is a tutorial for face recognition for Nano:

You can update the sample to face identification in a similar way.
Thanks.

Thank you. You are correct. I had read an article that face_recognition would not use GPU. Based on your feedback, I installed it, ran jtop, and it showed heavy GPU usage. Thank you for the help.