Computing landmarks transforms multithreading on CPU

[hqnicolas]

LandmarksHandler -> dlib.get_frontal_face_detector() -> alignment_utils.py -> get landmark with dlib
A numpy array, where each row contains a landmark/point x-y coordinates.
Return None if no face is detected by Dlib.

Do not have much experience with using ThreadPool
but I believe that its the bottleneck for sg3 inference in video
I've been passing files of 3000 frames
and the registration of the compute landmarks transforms take more time than inference

[yuval-alaluf]

This is probably possible to parallelize using multiprocessing. You are free to try this out.