Fusion Body-Tracking GPU load at 100%

Software
1

Hi,

I’m trying to use 4 ZED2 cameras connected to a strong PC with a Geforce RTX 4090 and using the python multi-camera example. Running this, the GPU is constantly at 100% GPU load even when setting the model to HUMAN_BODY_FAST or lowering the resolution or frame rate. Is this expected behavior? I remember reading in other posts that a 4090 should be powerful enough to handle 4 cameras.

Kind regards,
Moritz

2

Hi,

For multi-camera application, I’d highly recommend using C++ instead of Python.

Because of the Python Global Interpreter Lock (GIL), multi threaded applications are not very efficient in this programing language compared to C++.

3

Hi Benjamin,

thanks for the quick answer. Do you have any experience regarding the performance (4 cameras, 1 4090) with body tracking or object detection on python and c++ systems? Just so we know what we can realistically expect?

And shouldn’t the CPU be more of a problem than the GPU if multi-threading were the issue?

In any case, I’d have assumed that a 4090 can easily handle 4 HUMAN_BODY_FAST models if one Jetson Nano can handle it or am I thinking wrong here?

4

Any input here would be greatly appreciated :slight_smile:

5

Hi,

I don’t have the number for a 4090, but with a 4070 Ti, you can expect around 12/15fps for body tracking depending on the body model (18 or 38) with Neural Light depth mode.

Stereolabs Support

6

Hello Benjamin,

thanks for the feedback! Just to be on the safe side: With four cameras we can expect 12/15 fps, with one we could get the FPS of the camera?

7

yes, with one camera, you can expect 60fps.

Stereolabs Support