Replace opencv with gstreamer

[roiksail]

How to change this code with opencv instead of gstreamer?
Whats difference of between gstreamer and opencv in cpu usage and speed?
My goal for this project is to have a large number of cameras for high fps processing.
For example 20 cameras
But the problem is that more than 7 cameras,my cpu usage reaches over 98% and it is not possible to add a new one.
How can I do this project on 20 cameras at a time?
490mb is model loaded by gpu for simgle camera and I probably won't get down to the gpu resource source. I think my problem is the cpu resource
gpu: 1080ti(11gb), cpu: core i7 9700k
Or at least how to choose the system I need for this project?

[TorRient]

Same question :D
Faster-mobile-retinaface can't detect landmark, right?

[1996scarlet]

How to change this code with opencv instead of gstreamer?
Whats difference of between gstreamer and opencv in cpu usage and speed?
My goal for this project is to have a large number of cameras for high fps processing.
For example 20 cameras
But the problem is that more than 7 cameras,my cpu usage reaches over 98% and it is not possible to add a new one.
How can I do this project on 20 cameras at a time?
490mb is model loaded by gpu for simgle camera and I probably won't get down to the gpu resource source. I think my problem is the cpu resource
gpu: 1080ti(11gb), cpu: core i7 9700k
Or at least how to choose the system I need for this project?

Decoding the H.264 (or other format) stream using CPU can cost much.
I'd suggest using your NVIDIA GPU for decoding acceleration.
By the way, opencv's default decoding implementation is slower than using gstreamer or ffmpeg.

Here is a GPU decoding example using ffmpeg with Nvidia-P40, (20 cameras are piece of cakes, open 20 scripts or what you like）

import os
import cv2
import numpy as np
import subprocess as sp
import time
import sys
 
V_W, V_H, V_C = 1920, 1080, 3
BUFFER_SIZE = V_W * V_H * V_C
 
cmd_in = ['ffmpeg',  '-hwaccel_device', '0', '-hwaccel', 'cuvid',
'-c:v', 'h264_cuvid', '-resize', f'{V_W}x{V_H}', '-vsync', '1',
'-i', f'RSTP ADDRESS',
'-vf', 'scale_npp=format=yuv420p,hwdownload,format=yuv420p',
'-f', 'image2pipe', '-pix_fmt', 'bgr24', '-vcodec', 'rawvideo', '-']
 
pipe_in = sp.Popen(cmd_in, stdout=sp.PIPE)
 
# cmd_out = ['ffmpeg', '-y', '-f', 'rawvideo', '-vcodec', 'rawvideo',
#            '-pix_fmt', 'bgr24', '-s', f'{V_W}x{V_H}', '-i', '-', '-c:v',
#            'h264_nvenc', '-pix_fmt', 'yuv420p', '-preset', 'fast',
#            '-f', 'flv', f'rtmp://127.0.0.1:1935/hk{index}/livestream']
 
# pipe_out = sp.Popen(cmd_out, stdin=sp.PIPE)
 
while True:
    src = pipe_in.stdout.read(BUFFER_SIZE)
    frame = np.frombuffer(src[:BUFFER_SIZE], dtype=np.uint8)
    frame = frame_22.reshape(V_H, V_W, V_C)
    # pipe_out.stdin.write(frame.tostring())
    
    cv2.imshow('res', frame)
    if cv2.waitKey(10) == ord('q'):
        break

[1996scarlet]

Same question :D
Faster-mobile-retinaface can't detect landmark, right?

lol, I removed the landmark branch.

[roiksail]

Thank you
I do one camera in a single core process (multiprocessing)
And for this one camera I load the retinaface model once.
So when I have 20 cameras I have to load this model 20 times for each camera.
So when each model gets 1500 megabytes of graphics card then I get about 20 cameras ,I need 30gigabytes of graphics memory, which is not economical at all.
Am I going the right way?How do I do this on multiple cameras at the same time in the most effective way possible?
I agree that ffmpeg can be a lot more effective than opencv, but with the system I mentioned above how should I implement this system?
Do we need to have multiple separate Python files(scripts) and run them at the same time? Or is there a better way to do this?

[bao-O]

@roiksail Did you successfully run this repo with opencv? And if yes, how about the fps when comparing with gstreamer? Plz shared your work with me if you already done it.

[gjm441]

How to change this code with opencv instead of gstreamer?
Whats difference of between gstreamer and opencv in cpu usage and speed?
My goal for this project is to have a large number of cameras for high fps processing.
For example 20 cameras
But the problem is that more than 7 cameras,my cpu usage reaches over 98% and it is not possible to add a new one.
How can I do this project on 20 cameras at a time?
490mb is model loaded by gpu for simgle camera and I probably won't get down to the gpu resource source. I think my problem is the cpu resource
gpu: 1080ti(11gb), cpu: core i7 9700k
Or at least how to choose the system I need for this project?

Decoding the H.264 (or other format) stream using CPU can cost much.
I'd suggest using your NVIDIA GPU for decoding acceleration.
By the way, opencv's default decoding implementation is slower than using gstreamer or ffmpeg.

Here is a GPU decoding example using ffmpeg with Nvidia-P40, (20 cameras are piece of cakes, open 20 scripts or what you like）

import os
import cv2
import numpy as np
import subprocess as sp
import time
import sys
 
V_W, V_H, V_C = 1920, 1080, 3
BUFFER_SIZE = V_W * V_H * V_C
 
cmd_in = ['ffmpeg',  '-hwaccel_device', '0', '-hwaccel', 'cuvid',
'-c:v', 'h264_cuvid', '-resize', f'{V_W}x{V_H}', '-vsync', '1',
'-i', f'RSTP ADDRESS',
'-vf', 'scale_npp=format=yuv420p,hwdownload,format=yuv420p',
'-f', 'image2pipe', '-pix_fmt', 'bgr24', '-vcodec', 'rawvideo', '-']
 
pipe_in = sp.Popen(cmd_in, stdout=sp.PIPE)
 
# cmd_out = ['ffmpeg', '-y', '-f', 'rawvideo', '-vcodec', 'rawvideo',
#            '-pix_fmt', 'bgr24', '-s', f'{V_W}x{V_H}', '-i', '-', '-c:v',
#            'h264_nvenc', '-pix_fmt', 'yuv420p', '-preset', 'fast',
#            '-f', 'flv', f'rtmp://127.0.0.1:1935/hk{index}/livestream']
 
# pipe_out = sp.Popen(cmd_out, stdin=sp.PIPE)
 
while True:
    src = pipe_in.stdout.read(BUFFER_SIZE)
    frame = np.frombuffer(src[:BUFFER_SIZE], dtype=np.uint8)
    frame = frame_22.reshape(V_H, V_W, V_C)
    # pipe_out.stdin.write(frame.tostring())
    
    cv2.imshow('res', frame)
    if cv2.waitKey(10) == ord('q'):
        break

Great work for sharing.
Do you remove the landmark branch by retrain network? or can edit params and jsons by scripts directly?