Very poor Performance with with NVIDIA Jetson Nano 2GB in Face Recognition

Autonomous Machines Jetson & Embedded Systems Jetson Nano
1

So i just started my adventure further into the AI and Deep Learning World. With a little help of @paul.mcwhorter Video Series on Youtube, and some additional stuff I read alongside. Now i’m at the point where the Facial recognition starts and i cant get a proper Frame rate for the Project it seems to be around .2 fps.

First of all some Information from my System.
NVIDIA Jetson Nano (Developer Kit Version)
L4T 32.6.1 [ JetPack 4.6 ]
Ubuntu 18.04.6 LTS
Kernel Version: 4.9.253-tegra
CUDA 10.2.300
CUDA Architecture: 5.3
OpenCV version: 4.1.1
OpenCV Cuda: NO
CUDNN: 8.2.1.32
TensorRT: 8.0.1.6
Vision Works: 1.6.0.501
VPI: ii libnvvpi1 1.1.15 arm64 NVIDIA Vision Programming Interface library
Vulcan: 1.2.70
dlib 19.17
face_recognition 1.3.0
Here is the Programm i try to run but it dont run properly:

import cv2
import pickle
import face_recognition
import os
#Variables Data
names = []
encodings = []
font = cv2.FONT_HERSHEY_SIMPLEX

#Load the Trainingfile
with open('train.pkl','rb') as file:
    names = pickle.load(file)
    encodings = pickle.load(file)

#Setup Cam 
cam = cv2.VideoCapture(0)
#Showing Cam
while True:
    #Looking for Faces in the Frame
    ret,frame = cam.read()
    frameSmall = cv2.resize(frame,(0,0),fx = .3, fy = .3)
    frameRGB = cv2.cvtColor(frameSmall, cv2.COLOR_BGR2RGB)
    facePosition = face_recognition.face_locations(frameRGB)
    allEncodings = face_recognition.face_encodings(frameRGB,facePosition)
    for (top,right,bottom,left), face_encoding in zip(facePosition,allEncodings):
        name = 'unknown Person'
        match = face_recognition.compare_faces(encodings,face_encoding)
        if True in match:
            first_match_index = match.index(True)
            name = names[first_match_index]
        top = top*3
        right = right*3
        bottom = bottom*3
        left = left*3
        cv2.rectangle(frame,(left,top),(right,bottom),(0,255,0),2)
        cv2.putText(frame, name,(left,top-6),font,.75,(0,0,255))
    #Showing The Frame
    cv2.imshow('Cam', frame)
    cv2.moveWindow('Cam',0,0)
    if cv2.waitKey(100) == ord('q'):
        break

cam.release()
cv2.destroyAllWindows

I appreciate every help

2

Thankyou for including your system information.

your issue (need for speed) is related to: OpenCV Cuda: NO
your using your CPU to process in cv2.

first check out.

also this thread:

3

Thanks for your answer. Sorry I didnt replied but I had a full schedule. I just upgraded my OpenCV and got CUDA Support in OpenCV. Its now looking like this:
NVIDIA Jetson Nano (Developer Kit Version)
L4T 32.6.1 [ JetPack 4.6 ]
Ubuntu 18.04.6 LTS
Kernel Version: 4.9.253-tegra
CUDA 10.2.300
CUDA Architecture: 5.3
OpenCV version: 4.5.5
OpenCV Cuda: YES
CUDNN: 8.2.1.32
TensorRT: 8.0.1.6
Vision Works: 1.6.0.501
VPI: ii libnvvpi1 1.1.15 arm64 NVIDIA Vision Programming Interface library
Vulcan: 1.2.70

However I build a FPS Counter to track how much FPS I get and it barely made it to 2 FPS which is horrible. My programm uses OpenCV 4.5.5 (checked it with cv2.version). I also checked in with jtop to see GPU usage when im running the Programm and it seems like the Programm still doesnt use any of it.

4

Is it not the face_recognition which is the problematic part since OpenCV just grabs the Frame draws some Circle and thats it. Nothing heavy load there or am I missing something. face_recognition compares the Faces so I think this is my Problem isnt it ?

5

I don’t see you using your GPU to process this in your python script.

net.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)
net.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA)

again please read this:

and this for more help:

Sorry im not a python guy or I’d be able to help you more. Want to do this inTensorFlow.js nodejs I’d be of more help 2 you.

1 Like
6

Yeah its okay I appreciate your help. But I dont understand I use the face_recognition library as my model if im right or ? So if i use the Lines you appended I will use another model am I right ? I read the other Forum post but i dont quiete get it.

7

Looks Like your right on the model needing to be different.

Check out this howto:

I took a look at his code provided for the example: https://techvidvan.s3.amazonaws.com/machine-learning-projects/face-detection-opencv-python.zip

It looks like you need to use the cnn model.
i found this in his example code.

# Find the face location and encodings for the current frame
    # 'cnn' runs on gpu and it is more accurate. change it to 'hog' is you want to run on cpu.
    face_locations = face_recognition.face_locations(Current_image, model='cnn')  
    face_encodes = face_recognition.face_encodings(Current_image,face_locations)

I hope this puts you on the right path.

Side Note:
If you start working on object detection using openCV and python let me know I’m working on that now.
I love nvidia and they have some great tutorials:

but they lock you into their eco system with their great tools. but having to import nvida-utils is not on my todo list

on day 12 of learning python and I’ve Just got openCV + cv2.cuda_GpuMat + python cropping the rtsp stream using the GPU.

code snippit:

while True:
            # read current frame
            _, img = camera.read()

            img_to_gpu = cv2.cuda_GpuMat()
            img_to_gpu.upload(img)

            #imgCrop = img[yStart:heigth_of_crop, xStart:width_of_crop] #CPU CROP
            #imgCrop = cv2.UMat(img, [yStart, heigth_of_crop], [xStart, width_of_crop]) #GPU UMAT CROP
            imgCrop = cv2.cuda_GpuMat(img_to_gpu, [yStart, heigth_of_crop], [xStart, width_of_crop]) #GPU cuda_UMAT CROP
            imgCrop = imgCrop.download()
            

            # encode as a jpeg image and return it
            yield cv2.imencode('.jpg', imgCrop)[1].tobytes()

if you head down the object detection route give me a PM (private msg] and I’ll see if I can help you out as that is what im working on now.

8

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.