Multi-camera mode

[MichaelBryce90]

Hey! You have a wonderful project. Tell me, if possible, how to run the example "Calculating the speed of cars using YOLO v4 in real time" and other examples in this repository in multi-camera mode. I want to try to run simultaneously recorded video and video from the ip camera in real time. Is it possible to run 3,5,10 video sources at the same time?

Thank you in advance!

P.S Please make a "Sponsor" button. I want to support your wonderful project! I think that many here are ready to reward your work!

[Nuzhny007]

Hi!
It need to write some advanced code. I can implement this function on weekend.

[MichaelBryce90]

Hi! It need to write some advanced code. I can implement this function on weekend.

Thank you for your prompt reply. I will look forward to! The project is very cool. I've been playing with examples for three days and I like everything. I also read all the articles on the basis of which the examples were created Thanks again!

[Nuzhny007]

I thought a little about the multicamera mode.
We have N cameras and M GPUs.

1. Now you can run N processes and they will be work parallel. And for each process set GPU index. Isn't that what you want?

2. Or I can create single process:

• M threads with YOLO (one thread per GPU with Detector) + and frames queues;
• N threads with Capture+Tracking.

What do you think about this? The (1) or (2)? Maybe anything else?

[MichaelBryce90]

I thought a little about the multicamera mode. We have N cameras and M GPUs.

1. Now you can run N processes and they will be work parallel. And for each process set GPU index. Isn't that what you want?
2. Or I can create single process:

• M threads with YOLO (one thread per GPU with Detector) + and frames queues;
• N threads with Capture+Tracking.

What do you think about this? The (1) or (2)? Maybe anything else?

Hello, Sergey! Thank you again for responding. As I understand it, the first option you suggested will work fine on a GPU and it seems to be preferable. I tried several times to run in parallel the same Darknet-based Yolo model on the 2080RTX TI and did not see any drawdowns.
And option number two would be more preferable for running on a CPU? Did I understand everything correctly? Regarding the second option, I have no arguments and experience. Will we get performance drawdowns when implementing option two?

[MichaelBryce90]

I thought a little about the multicamera mode. We have N cameras and M GPUs.

1. Now you can run N processes and they will be work parallel. And for each process set GPU index. Isn't that what you want?
2. Or I can create single process:

• M threads with YOLO (one thread per GPU with Detector) + and frames queues;
• N threads with Capture+Tracking.

What do you think about this? The (1) or (2)? Maybe anything else?

If we use the approach of the first option, apparently we can manage loads more flexibly with multiple GPUs, right? Or, to be precise, divide the processing and loads between the video cards, if there are several of them in the system.

[Nuzhny007]

The first approach you can use just now: it more robust but don't so flexible.
The second - like Triton server from Nvidia ( https://github.com/triton-inference-server/server ): multiple models, dynamic batching etc. It more complex and complicated. It can use one model for multiple cameras and add frames from cameras in one batch. To be honest, I don't really want to implement this.

[MichaelBryce90]

The first approach you can use just now: it more robust but don't so flexible. The second - like Triton server from Nvidia ( https://github.com/triton-inference-server/server ): multiple models, dynamic batching etc. It more complex and complicated. It can use one model for multiple cameras and add frames from cameras in one batch. To be honest, I don't really want to implement this.

Sergey, hello! I apologize for the long answer! Thank you very much for the multi-camera implementation. I'm definitely going to try it next weekend!