Faster face detection on video stream

• face_recognition version: 1.2.2
• Python version: 3.5
• Operating System: Ubuntu 16.04

Description

Is there any way to make the face detection part run faster?
I'm running this for my rtsp camera stream and the video is very slow - each frame takes 2 seconds to process.
I managed to accelerate my video stream to 3 FPS by using "cnn" face detection method, by resizing the processing frame by a factor of 0.33 and by moving all the processing function calls into the separate thread.

I measured time it takes to process one frame - and it is somewhere around 350 ms.
But this is done in the separate thread (not the one that handles video display), so I don't understand why does this slow my video so much.

I am running this program on Nvidia Jetson TX1.

What I Did

# The main cycle part
for face_location, name in zip(face_locations, face_names):
    top, right, bottom, left = face_location
    cv2.rectangle(img, (left * 3, top * 3), (right * 3, bottom * 3), (0, 0, 230), 3)
if obama_available:
    oth = Thread(target=obama_thread, args=[copy.deepcopy(img)])
    oth.daemon = True
    oth.start()

## in the other thread
def obama_thread(img):
    obama_available = False
    small_frame = cv2.resize(img, (0, 0), fx=0.33, fy=0.33)
    rgb_small_frame = small_frame[:, :, ::-1]
    face_locations_temp = face_recognition.face_locations(rgb_small_frame, 1, "cnn")
    face_locations.clear()
    for (top, right, bottom, left) in face_locations_temp:
        face_locations.append([top, right, bottom, left])
    obama_available = True

Start by reproducing the RTSP workload on the Nvidia Jetson TX1 and measuring the call to face_recognition.face_locations in obama_thread, including the resize and thread handoff. Compare frame-display timing with the cnn detector at the reported scale; done should mean identifying the bottleneck and documenting a verified way to improve or explain the processing speed.