Very cool project.

I wonder if compressed sensing would be useful here. In the paper they describe this custom (and seemingly ad-hoc) video compression algorithm that is suitable for the low-power device. Basically, the camera averages blocks of pixels into "super pixels" all in the analog domain, sends these super pixels over the wireless link, and then the receiver can request the individual pixels of any super pixel that seems like it might have changed since the last frame. But with compressed sensing, can't all of this be done by sending random linear combinations of all pixel values (like the single pixel camera)? It seems like then you can avoid the bi-directional link and also get rate adaptation for free.