I think one possible explanation of why in CreateML is not possible to have an augmentation option in object detection model, is the fact that, under some circumstances, you would need to re-elaborate bounding boxes (in image classification you don't have bounding boxes locating objects). In some cases like color/saturation changes there should not be any re-elaboration of boxes and on other image transformations (e.g. flip) it could be very simple and not cpu intensive. But for streching, cropping and so on, it could rather difficult to re-annotate correctly the objects and may be cpu expensive... This is the answer I gave to myself... but I think, at the same time, I would like to have this option... because I'm doing Image augmentation by my self....and it's annoying!