search

GeekAlexis / FastMOT

High-performance multiple object tracking based on YOLO, Deep SORT, and KLT 🚀
MIT License
1.16k stars 255 forks source link

Making sure FastMot is running on GPU #155

Closed mortezafayazi closed 3 years ago

mortezafayazi commented 3 years ago

Hi,

I am running FastMot on a Jetson xavier device. Even though it works well, when I profile the GPU usage using jtop, it seems it is running on just CPU and GPU is not used. Although my FPS is around 5 however, it seems it should be around 15. It is possible that the code is running on CPU instead of GPU? If so how can I make sure it runs on GPU?

GeekAlexis commented 3 years ago

This is probably related to your installation/environment. You can search past issues on low FPS. Keep in mind that GPU is only used every 5 frames by default. So you should observe bursts of GPU usage. It’s not efficient to run everything on GPU.

mortezafayazi commented 3 years ago

Thanks a lot for your response. I checked the packages and here is their version: cuda: 10.2 cudnn: 8.0.0 tensorRT: 7.1.3 openCV: 4.5.1 numpy: 1.18.5 scipy: 1.5.0 Tenserflow: 1.15.5 Numba: 0.48.0

Also, here is how our GPU is used:

I get around an FPS around 10 for nyc1.mp4 video (using pyhon3 app.py --input_uri nyc1.mp4 --mot) that was uploaded here in issue #1. Do you have any suggestion why our FPS is too low?

Thanks a lot again for your time

GeekAlexis commented 3 years ago

GPU usage looks normal. It looks like you reinstalled OpenCV so marking this as a duplicate.