cleardusk/3DDFA

300W LP landmark issue

amyburden opened this issue · 4 comments

Thanks so much for providing dataset and model for us. I found some problems with the 300WLP dataset. Some landmark is wrong.
300W_LP/IBUG/IBUG_image_011_1_0.mat
300W_LP/IBUG/IBUG_image_011_1_10.mat
The files above have the exact same landmark while their faces are rotated. Besides, I would like to know what is the order of 300WLP pose annotation. The first three are pitch, yaw, and roll. They are in rad. How about f, x, y, z? Also, the angle seems to be the opposite. For example, positive pitch means looking down and in 300WLP is negative. positive yaw means looking left, while 300WLP is negative.

I have met the same problem. I found all rotated images has the same landmark. Is it normal?

I have met the same problem. I found all rotated images has the same landmark. Is it normal?

I have solve the problem. Do not use google search to download the 300 WLP. Use the link mentioned in the readme in this repo.
https://drive.google.com/file/d/17LfvBZFAeXt0ACPnVckfdrLTMHUpIQqE/view

I have met the same problem. I found all rotated images has the same landmark. Is it normal?

I have solve the problem. Do not use google search to download the 300 WLP. Use the link mentioned in the readme in this repo.
https://drive.google.com/file/d/17LfvBZFAeXt0ACPnVckfdrLTMHUpIQqE/view

Thanks for your reply. I have downloaded the links and got files "train_aug_120x120.zip, param_all_norm.pkl...", which only contain images and 3dmm params.

But as I want to do landmark research using 300W_LP dataset, the direct landmark ground truth is needed. It may be possible to use 3dmm ground truth to generate landmark points instead of using landmark from ".mat" file in the dataset.

I have met the same problem. I found all rotated images has the same landmark. Is it normal?

I have solve the problem. Do not use google search to download the 300 WLP. Use the link mentioned in the readme in this repo.
https://drive.google.com/file/d/17LfvBZFAeXt0ACPnVckfdrLTMHUpIQqE/view

Thanks for your reply. I have downloaded the links and got files "train_aug_120x120.zip, param_all_norm.pkl...", which only contain images and 3dmm params.

But as I want to do landmark research using 300W_LP dataset, the direct landmark ground truth is needed. It may be possible to use 3dmm ground truth to generate landmark points instead of using landmark from ".mat" file in the dataset.

hello.dude.Have you solved that?Can you give me some hints for that?