search

yfeng95 / PRNet

Joint 3D Face Reconstruction and Dense Alignment with Position Map Regression Network (ECCV 2018)
http://openaccess.thecvf.com/content_ECCV_2018/papers/Yao_Feng_Joint_3D_Face_ECCV_2018_paper.pdf
MIT License
4.95k stars 947 forks source link

About the speed #7

Closed mariolew closed 6 years ago

mariolew commented 6 years ago

Hi, nice work. I've tried the demo, however, I didn't get the speed reported in the paper, the average time to detect the test images is 100ms per image, my GPU is tesla K80 and I think the GPU is powerful enough.

yfeng95 commented 6 years ago

Hi, thanks for your interest in our work.

My test procedure are: Testing AFLW2000 images and only recording the time of the forwarding of PRN(as described in the paper), then calculate the mean time(ignoring the first two value). The result is, 9.8ms with a single 1080, I also tested it in other GPUs: ~10ms with TITAN X(pascal) and ~25ms with K40.

So, please make sure:

  1. your run time didn't include the detect(by dlib) and transform time. Specifically, please only print the time of prn.net_forward. (It seems you record the time of prn.process, which contains other parts like detection, cropping and transformation)
  2. If the time is still large, please comment other codes, only run net_forward in each iteration.(other codes have a little influence on the speed)
developer-mayuan commented 6 years ago

On 1080 Ti, I get around 0.01s for testing prn.net_forward. It's little wiered that 1080 is faster than TITAN xp.

petergerten commented 6 years ago

I am also curious about the 1080 vs TITAN X speed. Why is the TITAN X not faster?