About camera parameters

Hi, thanks for this nice work!

I am wondering that in case the cameras on different devices have different intrinsic parameters (like focal length), one same (or very similar) face appearance may lead to different PoGs. For example, if the focal length is not fixed, the distance from face to camera cannot be truly reflected by the appearance on the frame (in other words, the face box input to the model), hence such information will be confused. Is this something may happen?

Based on these thoughts, I'm wondering if the camera parameters in the dataset could be supported, then the frames could be re-projected into a fixed virtual camera.