Hello Arnage,
I fully understand your point, and assume that in some cases such approach could work.
But the problem is that there could be a lot of different capturing scenarios and cameras, so, for example, combination of landscape and portrait orientation can be used, also not every camera is writing the information about sensor rotation and so on.
However, if we find any solution that will provide "correct" coordinate system axis orientation in the most of cases without any referencing information, we will surely implement that.