Where is the source of ground truth redundancy in HD maps if they are compiled using the same neural network that does perception in real time?
Presumably it would be from statistics: if the last 1,000 times cars have passed through an intersection they have detected a stop sign, and now a car doesn’t detect a stop sign there, the car knows the HD map disagrees with its real time perception.
But what about when the environment changes? Some people say HD maps need to be updated daily because the built environment changes so frequently. When the HD map and real time perception NN disagree, how does the car know whether it’s a perception error or a change to the environment?
On the other hand, if the ground truth redundancy comes from human annotation, you’re going to need to collect raw sensor data on every mile of road, and update your HD maps more like once a year than once daily.
If the purpose of HD maps is something other than (or in addition to) sensing redundancy, what is it? Allowing the car to anticipate what’s up ahead on its route? But isn’t an HD map overkill for that?
I’m asking these questions 100% sincerely, not rhetorically — in case that isn’t clear.