How does Create ML prepare the images (MLImageClassifierBuilder)

Typically we start with a scaped data set. It would be nice to train off images we have collected, but in some cases thats too much work. So I have scraped imagaes of many different aspect ratios. My questing is.


Q) How does Creat ML prepare the images? I'm using MLImageClassifierBuilder.


The core algorithum uses a square. Are the images cropped, letterboxed or scaled?


My guess is they are scaled. The images shown in the GUI implies cropped but I think thats just the GUI. I guess scaling because typically you would have a single source for images and they would all be the same aspect ratio. Scalling uses all the image and the model is trained consistantly.


The reason I ask:


For my application I think I would be best off capturing a square image, that is, define my aspect ratio as 1:1.


Q) So to get the best from my scraped data set what should I do to them?


If I do need to map scaraped images to my desired 1:1 ratio should I be letterboxing (post boxing) or cropping?


Cropping could loose important parts of the image. Letter boxing adds unwanted "objects" into the scene and reduces the size of the object of ineterest.


I was thinking of using an intial trained model to search for badly cropped images and delete them from the data set and train again.


(The objects in the images are 3D solid objects so you would never see the object with a different aspect ratio. You will see it from different angles but they would look different, not just scalled.)