Landmark Prediction Stage(stage1)

[UdonDa]

Hi @YaN9-Y. Thank you for providing your codes.

I have two questions about the training of stage 1.

Firstly, According to the TRAIN_LANDMARK_LANDMARK_FLIST in config.yml.example, the value should be set to wflw_train_landmarks, ok.

However, in this setting, because the wflw landmark has the 196 points, should we set LANDMARK_POINTS to 98? The default value is set to 68 and you write the 68 in the paper.

Secondly, When we want to use the default value of 68, which landmark dataset should we use?

Thanks.

[YaN9-Y]

Hello Udonda, WFLW is a dataset for landmark detection task. We adopt this dataset to investigate whether the inpainting augmentation can help boosting the landmark prediction performance. If you want to hold experiments on WFLW dataset, the LANDMARK_POINTS should be set to 98. Otherwise, if you want to hold experiment on image inpainting, use celeba-hq dataset. The 68 points ground truth landmarks can be generated following steps in readme.

[UdonDa]

Thank you for replying.

So, if I use a celeba dataset, we set TRAIN_LANDMARK_LANDMARK_FLIST to celeba_train_landmarks.flist following this step(https://github.com/YaN9-Y/lafin#2-landmarks), right?

[YaN9-Y]

Right.

[UdonDa]

Thank you!