cnn-facial-landmark icon indicating copy to clipboard operation
cnn-facial-landmark copied to clipboard

Published 20 hours ago • verified publisher icon yinguobing

False prediction with cnn-facial-landmark ( own dataset )

Open abdou31 opened this issue 5 years ago • 2 comments

Hello Sir, I have trained my own dataset for eyes and iris detection from scratch, After reached more than 100000 steps , i get this error :

tensorflow.python.framework.errors_impl.NotFoundError: Failed to create a directory: ; No such file or directory So i think maybe is not a influential error , i used tensorfboard to see the graph : image

it seems good ( evolution avg loss ) . But when i tested , i get some bad results as you can see : iriss

Three things may causes this problem :

  • False annotation ( i annotated my dataset (300VW using OpenFace )
  • Starting from scratch ( pre-trained doesn't work , it causes problem )
  • I haven't 3D model for 40 landmarks

What can causes this false prediction?

abdou31 avatar Jun 05 '19 15:06 abdou31

Can you post some sample files you used for training?

yinguobing avatar Jun 11 '19 03:06 yinguobing

Did you mean .pts files ( annotation ) or what exactly do you talking about? If yes , this is an exemple of .pts file with 40 points ( points of the iris and the gaze extracted from openface csv output file using a python script ): image

abdou31 avatar Jun 11 '19 03:06 abdou31