Right, the "OCR" issue here is what I'm interested in --- Flock's vision models just straight-up making mistakes. There's a whole other issue of police not being well-trained in how to use these things, which leads to problems like "indict the first Dodge Durango we can find", but those are police policy issues more than they are ALPR issues.

▲

veec_cas_tant 14 hours ago | parent [-]

Does OCR accuracy even matter much here? I would expect, and maybe this is unrealistic, that anything even close to “this plate represents a stolen vehicle” would have at least one human involved. If I send in video evidence and someone just uses iOS OCR on a screenshot without checking it, I am definitely not blaming Apple.

	▲	tptacek 14 hours ago \| parent [-]
		In theory. Cards on the table, my prior on this is that these kinds of "OCR" errors are in fact very rare. I think that because we did a deep dive on every car curbed by a Flock camera here, as part of the process that led to us eliminating the cameras, and this just never happened. There is, at least in Illinois, a pervasive and difficult-to-solve accuracy problem with these systems, which is that Illinois LEADS is full of stale data; it wasn't designed with the expectation that it'd be matched against plates in real time, but rather with patrol cars calling in plates when they pulled them over.