Questions about data preprocessing

[jelleopard]

Hi Lixiong,

Thanks for your impressive work!
I have some questions on data preprocessing ways in your Faceptor, and hope I can get your answer.

In my understanding, landmark and face parsing use 512x512 as input, while others use non-aligned 112x112 images.
And I think facial bounding box is necessary to preprocess all the datasets.
So I wonder how to expand facial bounding box to match both two input sizes, or there are other preprocessing tricks in your paper?
And how to train the model with two different input sizes? (different from Swinface using 112 aligned images as input)

Thanks again for your time and hope to get your reply soon.