Can trt_pose run on dGPU?


I have run trt_pose project in jetson xavier AGX succeed, but when I set the resolution of video to 640x480 with resnet18, the perfomance is only 20fps, and the more high resolution, the more worse perfomance. (1920* 1080/1fps)

So I want deplore it to dGPU (T4), I wander whether it can run on dGPU? Thanks.


Suppose yes, but you will need to install all dependencies from scratch:

However, it looks like there are some issue on your benchmark result.
Since we can get 251 fps with resnet18 on Xavier.

Have you maximized the device performance first?

$ sudo nvpmodel -m 0
$ jetson_clocks


What resolution of your test? Maybe 224x224?
Have you try high resolution for detect?
And I noticed that if I want high resolution, I must recreate trt model so that can detect people.
If I dont rectreate trt model, I cant detect anything.


YES. We use the default 224x224 input.
For 640x480, the operation increase roughly 3x2 = 6 times.
So you should be able to get 251/6 ~ 40 fps, not 20 fps.

Just want to make sure again.
Do you maximize the device performance first?
More, which device do you use? Xavier 8GB, 16GB or 32GB?


I use xavier of 16GB, and set to 30W model.
I caculate the fps in execute() function, including prepocess and draw_object function.
I think it is more valueable rather than only caculate the infer performance.