(from email convo with Benjamin Elias Teitler) The API I'm using supports arbitrary images (so they don't have to be blocky QR codes), given that they have enough extractable features to distinguish from the rest of the room. It seems extremely robust and its lightning fast on both my iPod touch 4G, and my Galaxy S2.
https://ar.qualcomm.at/qdevnet/sdk
It was fairly easy for me to modify a sample to extract orientation. The catch is that they upload who knows what to their servers about your usage.
http://www.robots.ox.ac.uk/~bob/software/ptamm/manual.pdf has very detailed install instructions that might not be up to date. He recommends Linux all the way though…