Clarification on Landmark Scaling and Adjustments for Different Image Sizes

[bejujEbtNfubGvtjp]

Hello ! First of all, congratulations for such amazing work so far.

I have a few questions regarding the logic used for adapting landmarks to different image sizes in the following code snippet:

src = np.array([
    [30.2946, 51.6963],
    [65.5318, 51.5014],
    [48.0252, 71.7366],
    [33.5493, 92.3655],
    [62.7299, 92.2041]
], dtype=np.float32)

if image_size[1] == 112:
    src[:, 0] += 8.0

I have come to understand that this adjustment (src[:, 0] += 8.0) is made when resizing for images of size 112x112, as opposed to the default size of 112x96. However, I have the following questions:

1. Why is the horizontal adjustment specifically +8.0? Since 96 + 8 = 104, it doesn’t fully align with 112. Could you clarify the reasoning behind this choice?
2. If I need to adapt this logic for images of size 512x512, how should the src values and adjustments be scaled? Is it simply a linear scaling by a factor of 512 X 512​, or is there additional logic to consider?
3. How were the original src values determined? Were they derived from statistical averages of keypoints across a specific dataset, or do they originate from a standard alignment template? I would greatly appreciate an explanation of the mathematical principles involved. Additionally, if possible, could you provide a reference or source for this information?

Any clarification or guidance on these points would be greatly appreciated. Thank you for your time and for maintaining this repository!

[gremlation]

Why is the horizontal adjustment specifically +8.0? Since 96 + 8 = 104, it doesn’t fully align with 112.

Translating it right by 8 means that a 96-wide image is centered within 112 with padding of 8 on the left and padding of 8 on the right. Or in other words, left(8) + image(96) + right(8) = 112.

[bejujEbtNfubGvtjp]

Why is the horizontal adjustment specifically +8.0? Since 96 + 8 = 104, it doesn’t fully align with 112.

Translating it right by 8 means that a 96-wide image is centered within 112 with padding of 8 on the left and padding of 8 on the right. Or in other words, left(8) + image(96) + right(8) = 112.

Okay, I get it now. Thank you so much.

[bejujEbtNfubGvtjp]

I would really appreciate if someone could answer the remainder of the queries. Thank You !