Hi Hideo,

Thanks for asking; this is my answer:

1. Yes, it would help if you had ArcGIS Pro, the version I used previously 2.5, and the Image Analyst extension.
2. Your question is very conceptual. To make that easy, the cutting process is handled by the ArcGIS, so there will be no cross footprint across the tiles. The good part of using ArcGIS the footprint information extracted together with the tiles automatically, so you do need to put any effort there.
3. I believe your third question is about how object detection in deep learning works. Basically, because you train a model with references to “perfect” rooftops, the deep learning model will understand how to avoid conditions when one footprint detected as two or more footprints. Remember why we cut the image into tiles because we need to resize the image to reduce computing power. We also need to put some “knowledge” when we cut the building by making sure each tile consists of “perfect” rooftops. If you set the tile size too small it is very likely that your model will detect one footprint as several footprints.

To answer all your answers, just use ArcGIS Pro because it handles almost all technical issues you ask. We are not doing that manually.

Hope this will help you.

Best,
Lucas