interpolating protocol in image translation

[Jonathan Campeggio]

I am implementing some data augmentation to better the accuracy of an Image Classifier. So I use a series of Tensorflow preprocessing layers. In the specific case, I am using tensorflow.keras.layers.RandomTranslation, here you can find the documentation. I want to randomly translate the images. When I translate, there will be some points that will go outside the border. So I understand the argument fill_mode. I use the reflect mode, so for the pixels outside the boundaries, it will uses a reflective boundary conditions. So, why do I specify an interpolation mode? Interpolation is used to determined new points, but in the translational case, I have the original pixels, a translation vector, and an algorithm to fill the points outside the boundaries. Why do it need an interpolation mode?

I am citing Tensorflow as an example, but the question is quite simple and conceptual. Do I need to specify an interpolation protocol if I just translate an image?