The calculation method of scaling factors in the dataset labels.

[MambaKyrie]

I would like to inquire about the calculation method for scaling factors in the dataset labels because I am building my own dataset, but I have tried multiple times without success in finding the formula or approach for computing the scaling factors.

[gabriel-seegerer]

I'm not sure about the original method, but I found the scale value is closely related to the height of the bounding box created by the min/max of the ground truth landmarks.
As far as I can tell, scale is used to scale a region of scale * 200px (estimated face size) around the center_w, center_h values to the input size of 256px (see GetCropMatrix.process()).
My estimation is as follows:

tsv_path = osp.abspath('./data/datasets/annot/COFW/test.tsv')
with open(tsv_path, 'r') as f:
lines = f.readlines()
for line in lines:
items = line.strip().split('\t')
_, _, landmarks_gt, _, _, _ = items[:6]
lm = np.array(list(map(float, landmarks_gt.split(',')))).reshape(-1, 2)
bbox_height = lm[:, 1].max() - lm[:, 1].min()
scale = bbox_height / 200.0

In short:
Calculate your bbox height and divide by 200.

If compared to test.tsv of all datasets:
error: mean, max
300W: 0.0870, 2.0167
COFW: 0.0056, 0.0661
WFLW: 0.0305, 0.6162