' 0 - entity bounds ' 1 - region bounds ' 2 - region rect
0 will check if collision with rect bound of entire actor. The bound rect is never rotated so imagine that the collision area would be bigger if you had a rotated rectangular region in your actor.
1 will do the same as 0 but per region instead of the entire actor.
2 will test collision, per region and will use the regions rotation as well.
Currently there is no way to avoid white space detection unless you just split your image regions up by hand. So your "T" could be split into 3 separate regions for finer collision detection.