search

michalfaber / keras_Realtime_Multi-Person_Pose_Estimation

Keras version of Realtime Multi-Person Pose Estimation project
Other
779 stars 372 forks source link

About speed for testing one picture #58

Open MoonBunnyZZZ opened 6 years ago

MoonBunnyZZZ commented 6 years ago

I use the keras model to have a test on one jpg file,by 'python demo_image.py '.The processing takes 34 seconds.The jpg file is 500*500 pixels.May I consult you about how to use GPU to accelerate the testing @@process.

michalfaber commented 6 years ago

Hi @MoonBunnyZZZ What GPU card do you have?

MoonBunnyZZZ commented 6 years ago

@michalfaber GTX 1080TI

Uriya3 commented 6 years ago

How long the process should take on cpu i7 (windows 7, Intel 6th Gen, 16gb ddr4). On mine it takes more than 15 seconds per pic (low resolution)

IgorMunizS commented 6 years ago

My cpu i7 ubuntu 16gb ddr4 is taking around 25s in the sample image. Am I doing something wrong?

sungwonida commented 6 years ago

@Uriya3 @IgorMunizS Maybe because you're using CPU instead of GPU. In my case of using GTX1070 it can speed up to about x60 (compared to i7 CPU).

IgorMunizS commented 6 years ago

@sungwonida Thanks for your reply. It's true with GPU is much faster, but my doubt comes when I compare this code with Tensorflow implementation (https://github.com/ildoonet/tf-pose-estimation). Using same computer and same images I got ~16s. Both using CMU backbone. Maybe there is something different that I can't see.