At the bottom of the « collections » page in the technology paragraph, they mention confusing the Yolo object detector into thinking you’re a giraffe with medium confidence instead of a person
Yolo includes face detection, that's how it detects people. So it looks like yes this tech can be used for confounding more general image classifiers, but it was originally developed specifically for face detectors and the videos show it defeating those.
Specifically on cars, I don't think any of the currently deployed systems do face detection. Apart from anything else it takes too long, almost half a second on the fastest systems. They sense people as generalised blobs. It would be quite dangerous, to a system like that many advertising billboards would appear to be people sticking their faces right up against the camera.
A perfect one, sure, but in practice the car might, at best, treat it with less priority than a human in case of a difficult decision to make, and at worst, treat a « 40% confidence giraffe » as a false positive and ignore it
If a self driving car purposely ignores Giraffes or even real objects incorrectly identified as Giraffes or anything else, I think that's a problem. Africa is a place people drive cars in. And then there's this:
Specifically on cars, I don't think any of the currently deployed systems do face detection. Apart from anything else it takes too long, almost half a second on the fastest systems. They sense people as generalised blobs. It would be quite dangerous, to a system like that many advertising billboards would appear to be people sticking their faces right up against the camera.