If they’re seeing the code on the camera end it is very likely that the low confidence detections are not sent. The reason that you’d run local segmentation and detection is to reduce the amount of data being sent back. If they wanted maximum data from these devices they’d simply send the video and archive/process it all on the server end.

Image segmentation/classification models give false positives, it’s unavoidable. It isn’t evidence of anything other than the fallibility of computer vision technology.

I’m not saying that they’re not collecting other information about vehicles. We know that they also collect bluetooth identifiers and tire pressure monitoring sensors and other visual data about the car.

Only that the information in the article only shows that they’re doing segmentation and classification on the camera end. They don’t need “plausible deniability” as the original comment implies, there is nothing illegal about what they’re doing or anything that would leave them open to liability. (Obviously it’s creepy, immoral, and eroding personal freedoms… but I was only speaking of things that the Justice system cares about.)