I trained a CNN model for regression and getting facial landmarks of a face, it works well when input window has specific ratio but, it fails whenever I want to try it with bigger window and different scales and ratios. Now I want know how much is a CNN model robust to different input image ratios?