I don't think it would be feasible to do cloud analysis on a live video stream in self-driving context. The latency would simply be too high, and a single internet hiccup would result in a dangerous situation. So I think all the models run locally.