search

nikola1011 / yolov3-car-counter

This is a demo project that uses YoloV3 neural network to count vehicles on a given video
MIT License
7 stars 1 forks source link

it seems too slow #3

Open liuxufenfeiya opened 4 years ago

liuxufenfeiya commented 4 years ago

centos7,gpu p4000 8GB,i5 7500 i run this code,the effect seems great,the tracking is cool but my GPU memory is full,and the fps is less than 10 why did this so slow, other yolov3 detect code in my computer is more than 30 fps i read the code and find that every 10 imgs use yolov3 one time, so is the dlib's reason? result.png

nikola1011 commented 4 years ago

Hi @liuxufenfeiya,

The tensorflow is not the gpu version and no code should be running on GPU - propably that's why this is so "slow".

Please make sure you are not running anything else on your GPU.

liuxufenfeiya commented 4 years ago

试了cpu运行，更卡了，一张detect过程要0.4秒以上，远远大于gpu上的时间， dlib跟踪一辆车的时间不多，但是车一多，循环起来就很慢了，这也是为什么跟踪一张图要0.1秒的原因

nikola1011 commented 4 years ago

From translate.google.com: "Tried the CPU to run, and it got stuck. The detection process takes more than 0.4 seconds, which is much longer than the time on the GPU. dlib doesn't have much time to track a car, but when there are more cars, the cycle is very slow, which is why it takes 0.1 seconds to track a picture"

Thanks for checking it on CPU.

I looked at the code and until now, could not find a problem causing this. A more in depth investigation is required.