search

1adrianb / face-alignment

:fire: 2D and 3D Face alignment library build using pytorch
https://www.adrianbulat.com
BSD 3-Clause "New" or "Revised" License
6.94k stars 1.33k forks source link

low performance in Jetson nano #167

Closed SaddamBInSyed closed 4 years ago

SaddamBInSyed commented 4 years ago

Hi

thanks for your work.

I tested with jetson nano board. but the fps was very low.

is there any tuning needed to improve the same?

please advise

1adrianb commented 4 years ago

I am afraid I don't have access on a jetson board, but this is to be expected. Without retraining the network the simplest thing that you can do is to check which component is the slowest (face detection or keypoints localisation). Then, based on this adjust the resolution of the image, avoid running the detector on each frame, or replace it with a faster one. The other more involved options will require retraining the model using fewer stacks and fp16 or int8 precision.