More cameras for real time face recognition

[roiksail]

Hello everybody,im using arcface and mtcnn for real time face recognition and opencv for grab rtsp stream from my cameras,four ip cameras is processing for face recognition and i want to use more cameras for this project.
When i use four cameras my cpu usage is 90% and i couldnt to add more cameras.
How i can to use 20cameras for this project?

My system info is :gpu 1080ti,cpu core i7 9700k,32gb ram.
Isn't my hardware enough for this purpose?
How do I choose the right hardware?
Is retinaface a better target for my project?
How to run the whole process on gpu?
Please help me

[Luvata]

You should check the encode of your video stream, cv2.VideoCapture somehow use alot cpu

[roiksail]

Thank you so much,but how to encode rtsp live stream??
While i add my cameras ,face detection(retinaface) get 50% cpu usage.

[Luvata]

I used ffmpeg to capture the stream and the cpu usage is much lower (but I'm not sure if it can run real time). I haven't tried yet but you may try passing ffmpeg parameters in cv2.VideoCapture, or using ffmpeg-python

[Luvata]

@nttstar You can close this issue since there's no activity recently

[GanbI4]

Also, we met with mxnet's high cpu usage, converting to tvm helps us. We was able to achieve about 1 core per camera (1080p@25fps). Other possible optimizations:

1. Move ur image manipulations to GPU (opencv C++ can do a lot with CUDA)
2. If you using python, take look at cpu profile.

[luan1412167]

@GanbI4 Can you share your experiment for running on tvm? What is your model you using for face detection? Because I use R50 retinaface run on tvm(gpu gtx 2070) with <15fps. How to tune retinaface run on gpu? Thanks for your sharing.

[GanbI4]

@luan1412167 Unfortunately, i can't share my code, cuz it's under NDA.
For general purpose, retinaface R50 is too much, try mobelinet from yangfly.
#669

For tuning, try to reduce this params:
https://github.com/deepinsight/insightface/blob/master/RetinaFace/test.py#L10

About tvm, this tutorial helped me a lot:
https://docs.tvm.ai/tutorials/frontend/from_mxnet.html

[ahmetybilgin]

@Luvata @roiksail @luan1412167 Any progress in your work? Thanks.