ZhenglinZhou / STAR

[CVPR 2023] STAR Loss: Reducing Semantic Ambiguity in Facial Landmark Detection
168 stars 20 forks source link

Question about the data augmentation #27

Closed Yanyirong closed 8 months ago

Yanyirong commented 8 months ago

Hi @ZhenglinZhou :

Here is a question of the data augmentation.

In the AlignmentDataset the raw landmarks of a picture will be loaded and sent to do the augmentation. https://github.com/ZhenglinZhou/STAR/blob/2196c576bab0e88b8726e595437332f439cdfc1e/lib/dataset/alignmentDataset.py#L293

In the augmentation steps, the lmk will be transposed to aug_lmk https://github.com/ZhenglinZhou/STAR/blob/2196c576bab0e88b8726e595437332f439cdfc1e/lib/dataset/augmentation.py#L56

Then the augmented lmk will be sent to _norm_points. https://github.com/ZhenglinZhou/STAR/blob/2196c576bab0e88b8726e595437332f439cdfc1e/lib/dataset/alignmentDataset.py#L296

The the processed lmks will served as the labels in training steps.

So I want to know what is the mathematical meaning of some of the transformations done on the landmarks in this step of data augmentation. What format will the initial lmks be converted into for training?

I am new in this field and not familiar with some operations. Hope to get your help, thanks!!