I get the use of maps in the short term. Depending on the level of development, it is replacing a bigger problem (greater inaccuracy) with a smaller problem (lower inaccuracy). However, if the goal is to eventually minimize inaccuracy, then even that better option may not be best.
The basic issue I see with map as ground truth:
Are you, and the car, assuming the map is always correct?
If the map is fallible, how does the vehicle determine that?
In the case of city AP, yes Tesla is using the map as a gating function on stop light detection.
@Bladerskb , I was wrong in how it was currently operating. However, they are not gating the detection algorithm, just the resultant action. Assuming they can get the detection level up, I would expect the gating to be removed. However, I would also expect that they may retain the map as a fail safe, resulting in the union of the false positives.
Example why this is good: driving in an area with a blackout. No traffic indicators anywhere. Very helpful to know where intersections are. This is like the difference between a tourist vs a native to the area driver.