US military warboffins claim that they have found one of the internet's holiest grails – that of true, tagless searchable-video technology. For most of us, searching a video archive is a chancy matter. The perfect piece of footage we want to see may well be in there, but the only chance of finding it is usually that a person has …
A plausuble source
VIRAT will be a plausible source for much intelligence that really came from other sources that should not be compromised. It will work like this:
1. get intelligence from operative or cracked encryption
2. find a reason to fly a drone or "to have" had a statellite looking in the right place
3. pretend that VIRAT found the intelligence for you.
4. act on intelligence and hope the real source isn't suspected by anyone
interesting update to technology
sounds like a useful addition to existing software that is already looking at video from drones.
I remember hearing around 3 years ago about software they were using on video feeds from Afghanistan that was looking for changes in traffic (vehicles or human) i.e. a normally busy road or track suddenly stops being used by the locals or suddenly all of the traffic is sticking to one side of the road etc.
what they found was that in most cases the reason for the sudden change in behavior that the software flagged was that some form of IED had been placed on that road.
Sometime known as the Constance Babbington Smith ploy
CCTV auto recognition
Already exists. Mark Thomas went to Newham, where they use a system called Mandrake. It (is supposed to) recognise individuals from CCTV footage...
However it seems to work about as well as most DARPA projects :)
I'm not sure a lot of drone video is even *up* to CCTV quality.
El Reg reported the FBI shut down their facial recognition project running since the early 1960s (1962?) as the false recognition rate of *all* systems was still too high. Facial should be *much* simpler.
Casinos (who have a *very* keen interest in knowing who's *really* in their building) reckoned (IIRC) to be about 56% accurate (but that figures old).
It's impressive but suspicious.
can't wait to get my hands on this
Why pay some unspecified amount and have it be proprietary?
Why not keep enhancing motion tracking, pattern analysis, and keep adding more models of known objects to the database?
Teach the database to recognise original and natural modifications of things. Integrate all that with realtime intel feeds. What is Lockheed able to do that some well-rewarded, sufficiently-informed geeks on government payroll cannot? Oh... we're talking big money...
Which begs the possibility...
so it's a scene analysis/decription generating system
A fairly popular subject for AI research proposals (often bank rolled by those fine ladies and gentlemen at DARPA) since the 1970s.
Looks like the Nth generation go round *might* have finally proved flexible enough for deployment IRL.
The implementation project sounds like a classic LockMart deal.
It'll cost a fortune and *might* work.
But if it doesn't they'll just hide it behind "National security" and re-direct/re-scope to further research.
And bank the profits as per SOP.
Maybe scanning the video is not the problem
As others commented the ability to scan videos exists (your mileage will vary). So maybe part of this tool is an ability to take an instruction.
Some of the contract information refers to improving the query aspect. How do you describe to software the thing you want it to find?
If you've ever tried finding a music track by humming a few notes you'll know how difficult it is. We're OK at translating verbal instructions to, say, find something but far from perfect.