[londons_explore]

This is a perfect use for image to image ML models...

Throughout one recording, the phase shift caused by the distortion of the glass screen is probably approximately the same - and therefore could be learned.

Then for the actual decoding, certain elements of the frame should be of approximately known colours - for example someone's face should be skin colour. That then informs the colours for neighbouring objects, since over a small area phase is consistent.

Applying such techniques repeatedly over the whole video, trying to minimize inconsistencies, I'd bet you can get really good results.