ZhenglinZhou
commented
1 year ago Hi @yangli-lab, thanks for your interest!
The input params for preprocess function are img, landmarks_target, scale, center_w, center_h (see here, landmarks_5pts is not used in our exp).
The scale, center_w and center_h are associated with the detected face box and could be calculated as below:
- scale = min(x2 – x1, y2 – y1) / 200 * 1.05
- center_x = (x1 + x2) / 2
- center_y = (y1 + y2) / 2
Therefore, I suggest to first use a face detector to preprocess your own face images, and calculate the needed params. Then you could organize them in a tsv format (like this) or rewrite the AlignmentDataset.
yangli-lab
Thanks for your excellent work. I want to train the STAR model on my own dataset, could give some advice on how to prepare the training data. Can I finetune the model on my own dataset, with only the face images and corresponding landmarks. I see that in your code, only images and landmarks are the datas used to train the STAR model.