Big image, small faces = lots of CPU time #featurerequest

[salamanders]

I tried throwing a 4k x 3k photo at it, and it took a long time. (many minutes)
I then tried slicing the photo into lots of overlapping 300 x 300 px squares, and it took much less time, even considering the overlap.

I think this makes sense, given that there is less searching for "large" faces before burrowing down to the details looking for smaller faces, but it seems very error prone to hand-craft the tiling. Is there a way to build this sort of optimization into the library when a photo exceeds a certain dimension?