search

hzh8311 / pyhowfar

pytorch version of face-alignment-training(ICCV2017)
Other
48 stars 12 forks source link

Cannot get reasonable landmarks on 2D face alignment #2

Open tyshiwo opened 6 years ago

tyshiwo commented 6 years ago

Hi,

We have tried your code on 3D face alignment, which works fine. However, when you change the 3D coordinates to the 2D coordinates while keeping the codes and networks the same, we cannot achieve reasonable landmarks and the results are randomly distributed. Any suggestions may be helpful. Thanks!

hzh8311 commented 6 years ago

Did you use --pointType 2D to conduct the 2D face alignment experiment？ If so, it should be comparable with the 3D face alignment.

tyshiwo commented 6 years ago

Hi,

Yes, we changed to '2D' but failed to converge. Have you ever trained the model for 2D face alignment? Can you provide the pre-trained model?

hzh8311 commented 6 years ago

Yes, I trained a 2D face alignment model, and get the validation curve below (left: 2D): validation curve. So it is not a problem to train 2D model.

tyshiwo commented 6 years ago

Hi,

Which dataset did you use for training? I have used 300W_LP, and changed the --pointType to 2D, and got the following log: 111 png

Is it normal or not? Thanks.