They don't need all the vision data -- likely what they're doing is running their beta self-driving features in the background all the time, and only send data back (1) when the driver's input differs dramatically from what the self-driver would have done itself, presumably because the driver saw something their software missed, and/or (2) when a driver gets in a crash (at which point they can check if their algorithm would have prevented the accident).

They just need to send back data from a few seconds before either of those scenarios to quickly accumulate a giant library of one-in-a-million edge cases they can test future algorithm tweaks against. I think it's a reasonable strategy.