I am new for OpenStreetMap, and noticed that the OpenStreetMap is allowed to visit the DigitalGloble Maps API or Bing Maps API, and for the OSM contributor can choose different map layer such like DigitalGlobe Premium and Bing aerial imagery in edit mode.
Thus, here is my question: is it possible/allowed to detect road and building via deep learning or something like that from satellites images?
You would need to read the specific terms of use. I imagine some are restricted to manual tracing. Note that Bing only allow some of its imagery to be used. In particularly they do not allow oblique views to be used.
Also, you need to be aware of the rules on mechanical imports, and I think this would be treated as a mechanical import unless it was only used to suggest the location of buildings that were then individually checked by eye.
In particular, buildings need to be checked to see if they are already mapped, as far as possible checked to see that they haven’t been demolished since the image was created, and account taken of local offsets between the datum for surrounding objects and that for the image (in particular to make sure they ended up in the correct relationship to already mapped roads and railways, and didn’t overlap other buildings.)
I assume you don’t really mean satellite images, as most of the images on which buildings are visible are take from aircraft. Most real satellite images are probably US federal government (NASA) public domain (in USA) material.