[josefx]

Most important road signs have rather distinct shapes, standardized sizes and are angled towards oncoming traffic. Having an object with known shape aligned almost perfectly towards the camera is basically the best case for many primitive object detection algorithms.

[mathgeek]

True, but it’s equally important that a self-driving car be able to recognize a stop sign that is bent from a previous accident and facing an arbitrary angle (as well as one that is angled towards the car’s lane but applies to a different road).

[MereInterest]

And stop signs that have been altered in some way. For example, rural stop signs that are peppered with holes from pot shots must still be recognized. Snowy stop signs with the bottom half obscured by accumulated drift. Signs with a non-red sticker reading “WAR” placed below the word “STOP”.

And that’s not even getting into cases where you conditionally act like there’s a stop sign. The city of Houghton,MI has major streets along the side of a hill, and minor streets going up and down the hill. Every winter, sand is put down for traction, and every spring it is cleaned away. If there’s a late-season snowstorm after the spring cleaning, cars going downhill on the minor streets physically cannot stop, so everybody on the major streets looks uphill before crossing.

Short of location-dependent fine-tuned models, I’m not sure how machine learning could replicate the logic of “if snowy in late spring, grant right-of-way to cars headed downhill”.