Once again, powerful pattern matching being dressed up as "AI"
weather to follow.
A trio of Stanford computer scientists have developed a deep learning model to geolocate Google Street View images, meaning it can figure out generally where a picture was taken just by looking at it. The software is said to work well enough to beat top players in GeoGuessr, a popular online location-guessing game. That's not …
> Certainly the label "AI" is overly broad and over hyped.
To quote myself from a couple of days ago:
AI research is a bit of a weird one: one of the old sayings was "if we've figured out how to do it, it isn't an AI question any more"
Years ago, we had no idea how to do this and get it to work, Machine Vision most definitely a topic for AI research.
Now we can see how to do it (well, *one* way to achieve the goal), oh, it is "just pattern matching" so referring to it as anything "AI" is over hyping!
Do not go into AI research if you ever hope to have people quietly applauding your results.
Yes. The Macsyma system which arose out of research on mathematical software in the AI group at MIT in the 60's was heralded as the harbinger of GAI being just around the corner, while simultaneously removing that kind of symbolic algebra from what anyone would consider AI. I think there are strong parallels to the current iteration of AI LLM's.
"Google Lens"
Not really. When I search for the name of some hot babe in an on-line photo, Lens invariably leads me to the vendor sites for her dress, shoes, handbag, jewelry, etc.
That's why I have found it more efficient to search on nude images.
If you ask the people who dressed up AI as something exclusively-futuristic about a century ago, their mind would be blown by this pattern recognition ability.
You can either keep up the chase of making AI always something inherently unachievable, which I think is pointless, or go with the current sense that AI is just a synonym for modern machine learning.
This post has been deleted by its author
Nice attempt at redirection. I'm far, very far from an expert in the field of intelligence but I have no doubt that intelligence does indeed make use of powerful pattern matching - something we humans can observe constantly. "Making use of" and "being" are two entirely different things though...
I chuckled, but then I thought about it. Licence plates are a large but known set; for instance it's unusual to have more than four characters together. So if you start with the set of all combos up to four characters, and then keep degrading the image quality by an amount, and throw in some angles and sunshine, you can create a training set where you 100% already know the correct answers.
This set could all be procedurally generated too. I know it's a big compute task but for the sort of TLAs that are interested in such things it's a) a one time cost, and b) well inside their budgets.
You then end up with a CSI "zoom and enhancifier' that can couple with contextual parameters (country of issue format, vehicle colour, etc) to work by pattern matching the indistinct input rather than actually 'enhancing the image'.
I am not an expert in any way in these topics, as is probably clear.
Training a neural net to recognise a character set is a perfectly good use: I'm sure I've got more than one textbook on Image Processing that has (a smaller version) of that as one of the worked examples[1].
The only bit of your idea that I'd caution over is the procedurally generated bit: you run the very real risk of training the thing to look for odd characteristics that come from the generation code.
The nice[2] thing about ANPR is that you can't lose: so long as the photo is taken of a vehicle on the public highway (hint: choose where to site the camera) any failure to spot a reg number means you've got someone you wag a finger at. Maybe just needs a squirt of water on the plate, or you've caught one of those twats who think using a "fun" font goes well with their personalised plate (or even worse, uses bad spacing and cut-up letter shapes to pretend he got the personalised plate of his dreams).
[1] I'd give a citation but I need something that will read all of the textbooks in my library and then let me ask a question without being able to remember a specific phrase or term used. Oh, and only give accurate answers or "dunno", so cuts out the 'obvious' candidates for that task.
[2] from the p.o.v. of the student using the text book, of course that is all I meant.
There was a paper a few years ago where it was shown to be possible to decipher obfuscated characters - and I think number plates may have been used as a case in point - based on only a few pixels per character. The reason was simple: Common photo blurring algorithms, where you select an area on an image with sensitive info and blur it, tend to resolve the same letters to the same set of pixels arranged the same way, every time. You only needed to build a look up table.
Elsewhere I believe there are techniques to accurately de-blur faces in videos by taking advantage of time interpolation to produce a sharper still image.
...and make a good guess, within 15 miles of the correct location, a lot of the time...
Still way ahead of my sainted mother, who was once helping navigate from the passenger's seat and described the next POI as "going under the Blue Road". Figure out where you are with that input ya clever so-and-so!
In the UK that would be going under a motorway. Which is probably sufficient from the context of the road you were on at the time and direction you were driving in.
However, this contextual information is really difficult, decades back working on an voice assistant (precursor to Siri) it was working out the context which those working on semantic analysis found most challenging, because it often required cross matching with other data streams.
On the OS Seventh Series, that's be in about a mile.
Couldn't beat one of those printed on linen. My father had a fair number.
I'll work on hacking this AI mode, I can make a lot of money selling pictures with "evidence" of Boris, Starmer, Biden, and Trump drinking beer in Russia and China. Will people think that it's "true" ... sure, hacking is the most "accurate" truth these days. If you disagree then I'll post a picture of me in the El Reg, UK offices - this will be a joke until I post it on Facebook.
I think if people were worried, they'd have already applied for their properties to be obscured.
Of course, you can still be located as the geography, terrain, flora, and your neighbour's house probably have enough similarities that anybody looking will be three hundred metres out rather than two hundred...
I was wandering around the place where I grew up on street view over the weekend. Some of the blurring must be automatic as I came across a weird one where a tree was blurred on one side of a little village green, but perfectly visible on the other side. Perhaps the AI decided that a tree looked too much like an advert for a competitor when seen from that angle?
Free? Really?
Not even an anonymous free trial of a couple goes? nah, not interested enough to sacrifice another free email addy ;-)
Handing over an email address that will be spammed is not "free" in my book, even if I can create new ones at will and then block or dispose of them. Even that takes time and so is a "cost" to me of my free time over and above the time I'm prepared to spend to see if it's of any interest.
Nah, man. You want instant gratification plus instant results and no need to create dumb accounts and such...
It's fun, it's free, it doesn't need a sign in, and it has kittens. I mean, what's not to like if you have a few minutes spare and nothing to do (or want to "look busy")?
Credit to the researchers. Image recognition in bulk - given just HOW many images are generated and no doubt collated for data mining, this is a useful capability.
I can actually think of some practical uses within my own organisation for this. Engineers working in the field? Point camera at equipment and get a report back on the history of work done to said equipment and/or add to that history.
Creepy uses are what they are, but TLAs have probably had equivalent capability for some time...
I watched the videos on this a while ago, and they admitted that their AI was often not recognising the areas, it was recognising the camera artefacts used to take the photos.
For instance bugs or scratches on the lens, broken pixels, specific aberrations, pieces of the street view car which get captured in each image in some countries. On top of that often the pictures were taken at a similar time in a particular area (because they're taken in series by car) it can also key on sky tones. It might get some things from foliage colour, but it's likely not recognising a leaf shape or tree shape.
It's still quite impressive, but you wouldn't be able to randomly take a photo with your phone and have it geolocated to within 15miles in general, this is very specific to the Street View dataset.
A perfect example of the way this sort of pattern recognition system works (how any neural net works): in ways that make a mortal man scratch his head and day "huh, didn't spot that".
And also a nice illustration of how these weirdities can be a Good Thing: all the people scared of geolocating their house from a random photo can relax.
An image lookup from a StreetView picture (presumably removed from context and with metadata stripped - does StreetView put metadata into saved images?) back into StreetView could well be a useful feature (maybe not to me or you right now, but...).