search

1adrianb / face-alignment-training

Training code for the networks described in "How far are we from solving the 2D & 3D Face Alignment problem? (and a dataset of 230,000 3D facial landmarks)" paper.
https://www.adrianbulat.com/face-alignment
Other
145 stars 48 forks source link

training set image and landmark annotation #7

Closed wjtan99 closed 7 years ago

wjtan99 commented 7 years ago

First of all, thanks a lot for your code.

I followed your instruction to prepare the data set for training. I downed the landmark file, there are 4 subdirectories called AFW, HELEN, IBUG, and LFPW.

I also followed the instruction on your homepage and downloaded the dataset LS3D-W, while your post says I should download 300W-LP. In the LS3D-W, there is no directory called 300W-LP, neither directories called the names in the landmarks directory. Where can I downland the data set 300W-LP?

Thanks a lot for your help.

1adrianb commented 7 years ago

Hi @wjtan99 you can download the 300W-LP dataset from here: http://www.cbsr.ia.ac.cn/users/xiangyuzhu/projects/3DDFA/main.htm

wjtan99 commented 7 years ago

I also saw in all image directories of LS3D-W, there are .t7 file for every image. Are those predicted landmarks or ground truth? Can I use them as training set? Thanks.

1adrianb commented 7 years ago

@wjtan99 yes you can use them for training. They were obtained using the 2D-to-3D stack starting from the ground truth bounding boxes. I think there a few (like very few) with some wrong labels but this should not affect your results since I used them in the current form for the evaluations done in the paper. I will sanitise them anyway toward the end of this month.