Last time I looked at this company they just dumped your uploads into an unauthenticated gcp bucket. They just ran your photo thru an llm and asked for its location at the time, and the founder was doing something very weird (in my opinion) with scraping Tinder profiles.
Back when GeoSpy was available for everyone to use, I did a test where I just uploaded an image that had a black background and white text saying the location of a place and a textual description.
GeoSpy told me that it was the place mentioned in the picture, with the textual description as evidence.
You upload photos to tinder, and tinder has rough data on location provided (distance to you) i believe, the photos were the photos people posted on their tinder profile.
Sure, an impressive bit of tech, but the potential for misuse is immense.
To mock their user reviews...
> “Graylark helped me find the person I'm stalking in under 20 minutes. This tool is unbelievable — a true game-changer for those with restraining orders like me who just want to get back at them for that court order."
Do you think criminalizing an activity will stop criminals from highly lucrative criminal activity without going to North Korea levels of societal control?
I'm not diminishing the ethics debate, but it's crazy to me how easy it was for two non-technical rich dudes in a garage to build Clearview AI (And before vibe-coding!):
1. scrape billions of faces from the internet
2. `git clone` any off the shelf facial-recognition repo
Am I the only one that finds it amusing that conpanies like Google and Facebook sent Clearview legal letters complaining about scraping data from their sites?
Unless someone posts a photo of the stolen car with the numberplate still on, how would you identify YOUR car that way? Its not like cars are unique pieces. Same for bikes or anything else...
You don't have to precisely identify it, you only need to narrow it down to a high likelihood of being your vehicle. Then you can verify the VIN in person.
It wouldn't be hard to narrow things down:
Year/make/model/trim/color/region/timeliness will narrow down to a very small if not already unique subset of vehicles. And on top of that vehicles may often have unique stickers, accessories, or scratches which can further strengthen the case. Flock e.g. uses this data in their vehicle identification algo.
Did I completely miss the technical aspect of this blog? They list an improvement but no details on how they achieved it. It sounds like a trained embedding model and a vector search. All told though this just reads as boring product talk.
The picture in the article shows what looks like keypoint matching (ie, SIFT, SURF, FAST) between the query picture and the database. This can give an exact location if a picture of the location exists in their database.
They contrasted this with their prior technique which is more of an image classifier that can identify general location from image features. This approach does not require their database to contain a picture of the exact location.
First things first, its entirely possible to geolocate using just visual markers.
A bunch of startups did it around 2018 (most got bought by facebook, ie mapillary) They work by extracting keypoints from pictures and building a massive point cloud of identifiable key points.
But
That picture they use with supposed keypoint matching is wrong. None of those keypoints are reliable feature descriptors. They all are on foliage, which changes depending on season and wind. Geolocating that picture accurately _automatically_ using features is next to impossible.
Now, they might have a vibe based matcher which does some basic spatial comparison, but I'm not sure how reliable they are, especially given a large search radius.
The other interesting question is, where did they get their data from? I'm pretty sure google spent a lot of time making it really difficult to train from street view (lord knows we've tried.)
Edit the demo here: https://geospy.ai/ is much more what I recognise a bog standard VPS system does. Note that the user is matching buildings. Thats far more reliable way to do feature matching.
>They all are on foliage, which changes depending on season and wind. Geolocating that picture accurately _automatically_ using features is next to impossible.
Seems plausible enough to me. The trees are evergreens in a place that doesn't get snow, and the keypoints are mostly grounded on stable parts of the trees (trunks or thick branches), which barring gale-force winds probably don't fluctuate all that much.
The part that gives me pause are the keypoints that map the hood of the car to the pavement, and the point on the far right that maps the ledge to the pavement. How can a system robust enough to map foliage also return such blatant false matches?
short answer is that they are similar enough features to match. think of them as homophones (ie words that sound the same but have different meanings) in language. You need context to be able to filter them out. (https://github.com/polygon-software/python-visual-odometry/b...)
> don't fluctuate all that much.
Over time that doesn't bear out. Good features are areas of high contrast with nice clearly defined edges (text is great, so are buildings). branches move, which means they create lots of diffrent features depending on the wind, even light wind. when we were building out maps, we filtered as much greenery out as possible
our goal was to build technology to safeguard American freedom and prosperity...
...America deserves more. While Silicon Valley hype centers around LLMs, AGI, and SSI, our focus remains on visual intelligence—understanding the world we see with our eyes, what we call Visual Super Intelligence
It's kinda shocking to me how people are so willing to give tools to government agencies to track, spy, find, dox, and identify fellow citizens.
I guess I grew up drinking the 'American culture is one of mistrust of government' cool-aide, rather than 'American government has deep pockets' fruit punch.
I'm not sure if it's just an evolution of the times, or an actual erosion of principals (since when? 9/11?)
A lot of people lack the ability to do or credibly threaten sufficient violence for that to be actionable. Serving all the details up to the cops, who have nearly infinite ability to threaten violence, on a silver platter "here's your open and shut auto theft case, now go pad your stats" is the more tractable solution.
that is not that either though or is it? i mean say i found my vehicle on some platform for sale and then located it with their service, now what? i call the cops i suppose, i dont see how this is much different to calling them once they agreed to meet somewhere.
Yes, the term is strykers. Which can refer to the person that does it, or the actual stolen car that has been legitimized.
The stryker will find or buy from somewhere, a pool of unissued VINs that don't flag anything in the state registration system and match various vehicles (Dodge Chargers, Kias, Hyundais). Then when someone comes with that vehicle, they will strike a new Vin plate. Sometimes if they buy the VINs it will come in a package with plates. From there it's possible to get the vehicle registered, most likely under someone else's name that has no idea and they will sell / rent the car with tag etc. Though sometimes they will just make a fake plate too and then steal a real plate, swap it with the fake plate and put the real plate on the stolen car and sell it like that. In some cases / states they can actually get a title reissued.
Boats are even better, but much smaller market, just look up coast guard plates on Amazon.
Usually in the sophisticated thieves, it's the case that they buy a VIN from a car that was exported and not recorded as such. They then get a new copy of the title for a car that is no longer in the country and can request new factory stamped vin parts such as the suspension pillar.
The car looks completely legitimate to your average person with matching VINs it's just there are now two cars in two different places
"law enforcement could quickly locate and recover the stolen vehicle"
"law enforcement agencies can achieve faster resolutions, greater efficiency, and better outcomes for vehicle theft cases"
Could and would are two very different things in America.
In most cases, the police would simply do nothing.
Facial recognition technology (see Facebook auto-detecting your friends when uploading a photo) has existed for decades. Why do the police still post photos of suspects asking the public in help identifying so-and-so?* Can't they cross-reference with the DMV database or even Facebook to see if there are any matches?
*Although these days they even stopped doing that, I've seen cases where they blurred out the suspects face and then asked the public in help identifying them. They do this to protect the criminal's identity. Sigh. I wish we could bring back name and shame.
I’d be lying if I said I never ripped photos from listings on a certain popular online auction platform when selling stuff online (before the ubiquity of digital cameras and smartphones)
Articles seem to only go back to around 2024, how about 1.0? What was major enough to finally reach 1.0? Would be great to have a more technical blog post about what kind of major breakthroughs were discovered while developing this since the first discovery of this in 2022.
>The New York State Department of Motor Vehicles (DMV) is urging New York City residents to exercise caution when purchasing cars online, as they have recently identified a surge in the sale of stolen vehicles.
>As of Oct. 15 (2024), the DMV reports that it has recovered 228 stolen vehicles amounting to a value of $6.35 million. Of the 228 vehicles recovered this year, 149 were purchased by an unsuspecting victim after seeing a post on Facebook Marketplace or a similar site.
That is 65% of recovered stolen vehicles in New York was sold through online marked places.
It's very common with bikes, it wouldn't surprise me if they do it with cars too. I found my own stolen bike for sale on Facebook but still wasn't able to recover it. They just use a stolen or anonymized Facebook account so you can't easily figure out who is selling it.
I once saw a story about someone who saw her bike on sale even before it was stolen. The thieves announce the bikes they see regularly in the same place in the street and only steal them if they have a potential buyer.
That way they only take the risk when they need to, they don't need any storage area, and if they are caught it is only for 1 bike, not tens or hundreds.
Nice idea. Maybe do it for bicycles which are often more unique/personalized. Also, how are they going to identify identical model/trim/color cars when the license plates have been removed or switched?
Sounds more like “vehicle recovery” for the repossession market first and foremost.
A repo investigator for the bank locates the target vehicle via owner’s social media, takes photo of the car, shoots it into GeoSpy, then ganks the car based on given locations in the owner’s photos. Pair it up with ALPR hits across a city from national ALPR networks (to help correlate home/business/work patterns) and… wellp, there you go!
Nothing like an OSINT company like Bellingcat hasn't done before, it's just that in those cases it was done on citizens belonging to adversary countries. It was just a matter of time, I guess.
Last time I looked at this company they just dumped your uploads into an unauthenticated gcp bucket. They just ran your photo thru an llm and asked for its location at the time, and the founder was doing something very weird (in my opinion) with scraping Tinder profiles.
https://x.com/i/status/1786030866214326651
Back when GeoSpy was available for everyone to use, I did a test where I just uploaded an image that had a black background and white text saying the location of a place and a textual description.
GeoSpy told me that it was the place mentioned in the picture, with the textual description as evidence.
Yes, I have done the same and also used it to prompt inject to give some info about the prompt, but I don’t have that data anymore.
1 reply →
What were the jpgs in tge thotDBSmash folder? Did the founder use that bucket to collect photos of women? Or was that geolocation material, too…
> Or was that geolocation material, too…
Yeah I think we've figured out what inspired him to build the tool. :/
1 reply →
You upload photos to tinder, and tinder has rough data on location provided (distance to you) i believe, the photos were the photos people posted on their tinder profile.
Wow — that is probably the most suspicious set of three details ever shared on this forum. Thank you for enlightening us!
Sure, an impressive bit of tech, but the potential for misuse is immense.
To mock their user reviews...
> “Graylark helped me find the person I'm stalking in under 20 minutes. This tool is unbelievable — a true game-changer for those with restraining orders like me who just want to get back at them for that court order."
It will never not be misused. These types of apps should be illegal
Do you think criminalizing an activity will stop criminals from highly lucrative criminal activity without going to North Korea levels of societal control?
12 replies →
[dead]
> the potential for misuse is immense.
Who do you think is "sponsoring" this ? /s
Reminds me of the Clearview AI controversy[0].
I'm not diminishing the ethics debate, but it's crazy to me how easy it was for two non-technical rich dudes in a garage to build Clearview AI (And before vibe-coding!):
It was just a matter of when.
[0] https://en.wikipedia.org/wiki/Clearview_AI#History
Concerned but given the use can it be stopped ?
Yes. If one knows that someone has their identifiable data without consent it is a problem.
While they are pictures on the internet it is one thing, when you gather them all and put a label with a number then it is problematic.
Remember that FaceApp to make you older, younger etc? Imagine how much data those guys collected?
I know someone who submitted the face of a member of my family without consent. You could not even complain without agreeing with the TOS first
1 reply →
Am I the only one that finds it amusing that conpanies like Google and Facebook sent Clearview legal letters complaining about scraping data from their sites?
Unless someone posts a photo of the stolen car with the numberplate still on, how would you identify YOUR car that way? Its not like cars are unique pieces. Same for bikes or anything else...
You don't have to precisely identify it, you only need to narrow it down to a high likelihood of being your vehicle. Then you can verify the VIN in person.
It wouldn't be hard to narrow things down:
Year/make/model/trim/color/region/timeliness will narrow down to a very small if not already unique subset of vehicles. And on top of that vehicles may often have unique stickers, accessories, or scratches which can further strengthen the case. Flock e.g. uses this data in their vehicle identification algo.
I have no photos of my vehicles to initiate a search this way. Am I an outlier and it’s normal for people to keep vehicle pictures handy?
14 replies →
"law enforcement could quickly locate and recover the stolen vehicle"
geospy.ai: the real technology seems to be that they invented the world's thinnest veil
Clearview seems even worse to me, but yeah this is a good candidate for second place on the list of undesirable businesses
Hah. Good point. Allow me to counter: the name alone.
Clearview could be eye drops. GeoSpy: well...
I think Flock beat them to it
Did I completely miss the technical aspect of this blog? They list an improvement but no details on how they achieved it. It sounds like a trained embedding model and a vector search. All told though this just reads as boring product talk.
The picture in the article shows what looks like keypoint matching (ie, SIFT, SURF, FAST) between the query picture and the database. This can give an exact location if a picture of the location exists in their database.
They contrasted this with their prior technique which is more of an image classifier that can identify general location from image features. This approach does not require their database to contain a picture of the exact location.
I don't think you missed anything, but I don't think it's intended to be a technical blog either.
Ok so this looks like bullshit.
First things first, its entirely possible to geolocate using just visual markers.
A bunch of startups did it around 2018 (most got bought by facebook, ie mapillary) They work by extracting keypoints from pictures and building a massive point cloud of identifiable key points.
But
That picture they use with supposed keypoint matching is wrong. None of those keypoints are reliable feature descriptors. They all are on foliage, which changes depending on season and wind. Geolocating that picture accurately _automatically_ using features is next to impossible.
Now, they might have a vibe based matcher which does some basic spatial comparison, but I'm not sure how reliable they are, especially given a large search radius.
The other interesting question is, where did they get their data from? I'm pretty sure google spent a lot of time making it really difficult to train from street view (lord knows we've tried.)
Edit the demo here: https://geospy.ai/ is much more what I recognise a bog standard VPS system does. Note that the user is matching buildings. Thats far more reliable way to do feature matching.
>They all are on foliage, which changes depending on season and wind. Geolocating that picture accurately _automatically_ using features is next to impossible.
Seems plausible enough to me. The trees are evergreens in a place that doesn't get snow, and the keypoints are mostly grounded on stable parts of the trees (trunks or thick branches), which barring gale-force winds probably don't fluctuate all that much.
The part that gives me pause are the keypoints that map the hood of the car to the pavement, and the point on the far right that maps the ledge to the pavement. How can a system robust enough to map foliage also return such blatant false matches?
> return such blatant false matches
long answer, have a try on this demo: https://docs.opencv.org/4.x/dc/dc3/tutorial_py_matcher.html
short answer is that they are similar enough features to match. think of them as homophones (ie words that sound the same but have different meanings) in language. You need context to be able to filter them out. (https://github.com/polygon-software/python-visual-odometry/b...)
> don't fluctuate all that much.
Over time that doesn't bear out. Good features are areas of high contrast with nice clearly defined edges (text is great, so are buildings). branches move, which means they create lots of diffrent features depending on the wind, even light wind. when we were building out maps, we filtered as much greenery out as possible
More from this company:
Why America’s Heroes Deserve the Most Advanced AI
our goal was to build technology to safeguard American freedom and prosperity...
...America deserves more. While Silicon Valley hype centers around LLMs, AGI, and SSI, our focus remains on visual intelligence—understanding the world we see with our eyes, what we call Visual Super Intelligence
https://geospy.ai/blog/why-america-s-heroes-deserve-the-visu...
'Americas heros' deserve oversight.
Services like this (Flock, etc.) should either be illegal or accessible to everyone.
It's kinda shocking to me how people are so willing to give tools to government agencies to track, spy, find, dox, and identify fellow citizens.
I guess I grew up drinking the 'American culture is one of mistrust of government' cool-aide, rather than 'American government has deep pockets' fruit punch.
I'm not sure if it's just an evolution of the times, or an actual erosion of principals (since when? 9/11?)
6 replies →
If a thief is selling your stolen car, would that not be very easy to locate by showing interest in buying it? Am I missing something?
A lot of people lack the ability to do or credibly threaten sufficient violence for that to be actionable. Serving all the details up to the cops, who have nearly infinite ability to threaten violence, on a silver platter "here's your open and shut auto theft case, now go pad your stats" is the more tractable solution.
that is not that either though or is it? i mean say i found my vehicle on some platform for sale and then located it with their service, now what? i call the cops i suppose, i dont see how this is much different to calling them once they agreed to meet somewhere.
Wondering how theives can sell a stolen car. Do they have fake paperwork?
Yes, the term is strykers. Which can refer to the person that does it, or the actual stolen car that has been legitimized.
The stryker will find or buy from somewhere, a pool of unissued VINs that don't flag anything in the state registration system and match various vehicles (Dodge Chargers, Kias, Hyundais). Then when someone comes with that vehicle, they will strike a new Vin plate. Sometimes if they buy the VINs it will come in a package with plates. From there it's possible to get the vehicle registered, most likely under someone else's name that has no idea and they will sell / rent the car with tag etc. Though sometimes they will just make a fake plate too and then steal a real plate, swap it with the fake plate and put the real plate on the stolen car and sell it like that. In some cases / states they can actually get a title reissued.
Boats are even better, but much smaller market, just look up coast guard plates on Amazon.
Stiker vs Styker, is regional.
For Reference: Striker Music https://www.youtube.com/watch?v=AaTxkD5JFpg
Usually in the sophisticated thieves, it's the case that they buy a VIN from a car that was exported and not recorded as such. They then get a new copy of the title for a car that is no longer in the country and can request new factory stamped vin parts such as the suspension pillar. The car looks completely legitimate to your average person with matching VINs it's just there are now two cars in two different places
4 replies →
So the only way to know that this has been done is to read the OBD2 VIN or check all the resaleable parts for VINs?
It sounds like this scam would only get discovered when you go to the dealer for service, perhaps.
2 replies →
You can buy a totaled car for cheap and use its VIN.
Are VINs not tied to the make, model, and year?
2 replies →
The easiest way is to sell it at a steep discount to a buyer willing to accept the lack of paperwork for a good deal.
In some cases the car is completely disassembled and the parts are sold.
from what I know they sell it to other criminals which use it to commit further crimes or ship it off the continent
most commonly it is used for drugs in canada since every case I hear about ends up in forensics
Further and further into hyper-surveillance hell
"law enforcement could quickly locate and recover the stolen vehicle"
"law enforcement agencies can achieve faster resolutions, greater efficiency, and better outcomes for vehicle theft cases"
Could and would are two very different things in America.
In most cases, the police would simply do nothing.
Facial recognition technology (see Facebook auto-detecting your friends when uploading a photo) has existed for decades. Why do the police still post photos of suspects asking the public in help identifying so-and-so?* Can't they cross-reference with the DMV database or even Facebook to see if there are any matches?
*Although these days they even stopped doing that, I've seen cases where they blurred out the suspects face and then asked the public in help identifying them. They do this to protect the criminal's identity. Sigh. I wish we could bring back name and shame.
> Can't they cross-reference with the DMV database or even Facebook to see if there are any matches?
I think there are laws that bar them from doing that.
[flagged]
But who takes the photo of the stolen car?
The thief, when selling it.
I’d be lying if I said I never ripped photos from listings on a certain popular online auction platform when selling stuff online (before the ubiquity of digital cameras and smartphones)
Looking back it used to be way less advanced than what they have now, makes me wonder how this compares with flock.
Even if this was a real application, it is wrong
Thieves will just ask their favorite chat bot to change the background of the photo.
So you can find the place where car thieves took it to take pictures, already knowing which city it was in from the ad. How useful is that!
Founder of GeoSpy here,
Thanks for the post, AMA for anyone into computer vision or AI.
:)
Articles seem to only go back to around 2024, how about 1.0? What was major enough to finally reach 1.0? Would be great to have a more technical blog post about what kind of major breakthroughs were discovered while developing this since the first discovery of this in 2022.
Why do you have a folder called “thotDBSmash”? What were you collecting there and why?
how much legal fund you prepare
Ha, the first photo is Alamo Square looking up Fulton.
> Thieves often post stolen vehicles for sale on platforms like Facebook Marketplace or Craigslist
That seems like a stretch. That wouldn't even make sense for them to do. Strange claim to make.
>The New York State Department of Motor Vehicles (DMV) is urging New York City residents to exercise caution when purchasing cars online, as they have recently identified a surge in the sale of stolen vehicles.
>As of Oct. 15 (2024), the DMV reports that it has recovered 228 stolen vehicles amounting to a value of $6.35 million. Of the 228 vehicles recovered this year, 149 were purchased by an unsuspecting victim after seeing a post on Facebook Marketplace or a similar site.
That is 65% of recovered stolen vehicles in New York was sold through online marked places.
https://www.silive.com/crime-safety/2024/10/stolen-vehicles-...
I am curious of
1) How many total stolen vehicles there were
And
2) If 65% of recovered ones being from Marketplace means only the low-hanging fruit were found.
1 reply →
It's very common with bikes, it wouldn't surprise me if they do it with cars too. I found my own stolen bike for sale on Facebook but still wasn't able to recover it. They just use a stolen or anonymized Facebook account so you can't easily figure out who is selling it.
I once saw a story about someone who saw her bike on sale even before it was stolen. The thieves announce the bikes they see regularly in the same place in the street and only steal them if they have a potential buyer.
That way they only take the risk when they need to, they don't need any storage area, and if they are caught it is only for 1 bike, not tens or hundreds.
5 replies →
When you or a friend show up to a public place to buy it, it’s usually pretty clear who’s selling it.
1 reply →
[flagged]
You’re speaking pretty definitely too. Got any citations to back up your claims?
7 replies →
"Try our demo!"
Okay then, thinks EB, mentally trying to decide which photos to try it with.
"Look here's a picture of a place, and here's a pin on a map that shows you where it is!"
Yeah, I can do that without AI.
TrashFuture recently did a show on this company with the guy from Blood Work. Unfortunately it it's behind a paywall, but it's a great episode.
https://www.youtube.com/watch?v=Vo0gualrJa4
Go to 5:30 for a taste
Nice idea. Maybe do it for bicycles which are often more unique/personalized. Also, how are they going to identify identical model/trim/color cars when the license plates have been removed or switched?
Sounds more like “vehicle recovery” for the repossession market first and foremost.
A repo investigator for the bank locates the target vehicle via owner’s social media, takes photo of the car, shoots it into GeoSpy, then ganks the car based on given locations in the owner’s photos. Pair it up with ALPR hits across a city from national ALPR networks (to help correlate home/business/work patterns) and… wellp, there you go!
You can even skip the investigator in a lot of places, thanks to Flock. Dystopian.
wouldn't it be cheaper to just buy your location from a data broker?
You can even skip the investigator in a lot of places, thanks to Flock. Dystopian.
Flock is an ALPR.
can we use this to finally prove the moon landing is a hoax!!!
I mean comparisons already exist showing matching landmarks between apollo photos and topography from orbiting satellites
This entire demo is just a surveillance state dog-whistle.
"It's used for car theft!" except the intended use is obviously target government buyers for tracking citizens.
Nothing like an OSINT company like Bellingcat hasn't done before, it's just that in those cases it was done on citizens belonging to adversary countries. It was just a matter of time, I guess.
What?
1 reply →