Since most of the images are not square, it might be more efficient use of pixels to use CNN with a rectangular input.
To do this, we will need to make all the inputs the same orientation (landscape).
Obviously, this should not be combined with rotations augmentations using angles distant from 0 or 180 degrees, since that would defeat the purpose of this.
Since most of the images are not square, it might be more efficient use of pixels to use CNN with a rectangular input.
To do this, we will need to make all the inputs the same orientation (landscape).
Obviously, this should not be combined with rotations augmentations using angles distant from 0 or 180 degrees, since that would defeat the purpose of this.