Questions about location of landmarks, normalised frame and bounding box about clandmark HOT 2 CLOSED

Hi @cat3333,

I am glad you asked, since now if anyone has the similar question, I can refer to this answer :)

1)The first set are landmarks coordinates in the original image, while the second set represents the landmark coordinates in the normalized frame. The format is s_i = [x_i; y_i], i.e. the first row are x-coordinates of landmarks, second row are y-coordinates and each column represents a landmark, they are numbered as in the images in the paper.

2. Yes, the landmarks are actually located just in the normalized frame. The coordinates in the original image are computed by back projection (based on coefficients of the affine transform computed during the normalized frame acquisition, i.e. the in-plane rotation + scale to a fixed size + translation of the origin).

3. see answer to 1). The only problem is the re-numbering of landmarks due to missing self-occluded ones. However, it is easy to compute the order based on the frontal view. In MATLAB it is as follows:

visible_subset = {
  [1 2 3       7 8 9       12 13 14       17 18 19    21] % -profile
  [1 2 3 4 5 6 7 8 9 10 11 12 13 14       17 18 19 20 21] % -half-profile
  [1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21] %  frontal
  [1 2 3 4 5 6 7 8 9 10 11       14 15 16 17 18 19 20 21] %  half-profile
  [      4 5 6     9 10 11       14 15 16    18 19 20 21] %  profile
};

views = {'-profile', '-half-profile', 'frontal', 'half-profile', 'profile'};

I will soon upload the learning scripts for the jointmv detector, where this problem is addressed.

4. The bounding box is easy to draw, since you need its coordinates anyway. To get the enlarged bounding box which is actually used to form the normalized frame, you can either use the facebox coordinates and bw_margin of clandmark class, or you can play with the getHinv() and getH() methods, which return the affine transformation matrix to get points to the normalized frame or back to the original image.

5. H and W are actually fixed in the model, parameter bw, use getBaseWindowSize() method to get the width and height of the normalized frame.

from clandmark.

Thanks for your quick reply and detailed solutions!!

They help me a lot! =)

from clandmark.