How many on-the-fly augmentations per image for a single-class segmentation mode [R]

I’m training a single-class segmentation model for large rectangular artwork placed on the floor and photographed from above. We have around 3,000 accurately masked original images taken by six different photographers. They are not the same height and do not hold the camera in exactly the same way, so the photos naturally vary in: roll pitch yaw camera distance object coverage in the frame centering and X/Y shift orientation perspective lighting The photos taken with flagship iPhone. I want to use on-the-fl