Which CelebA landmark TXT file to use in gen_landmark_48.py

[macsermkiat]

There're 2 files to choose CelebA

Is it the one with align & cropped?

The landmark locations in "list_landmarks_align_celeba.txt" are based on the coordinates of align&cropped images.

Also, don't we have to call gen_landmark_12 and gen_landmark_24 after Pnet and Rnet gen data?

[andbleo]

I think the answer is neither. It looks like gen_landmark_48.py expects an annotated file with 15 columns, and the CelebA files only have 11 columns.

If you look at the default filename for the annotation file in gen_landmark_48.py, it is "trainImageList.txt", which I have seen here: https://github.com/AITTSMD/MTCNN-Tensorflow/blob/master/prepare_data/trainImageList.txt

That does have 15 columns, but it isn't using the CelebA dataset like the README says. I'm not sure if that is the right file to use or not though.

[nvinhphuc]

You have to merge 2 files including bounding box file and landmarks file. The format is: 000001.jpg 4 column of bounding box 11 column of landmarks