I'm not convinced by the quality of this. I took some screenshots of street view, not including any icons, and it identified them as completely the wrong city. One of them included the name of the town on a bus stop, which it completely failed to identify, placing the picture across the county, also asserting that it contained featured that it definitely didn't, such as thatched rooves (all rooves in the image were normal slate). I would make trust it to get me in the correct area of the country, but that's about it
After a bit more testing, it could successfully identify Buckingham Palace and St Michaels Mount (however the location wasn't great), however a street overlooking a beach in Cornwall was marked as Wales, apparently including house numbers in Welsh and English (despite Welsh using English numerals). It seems to work somewhat ok if there is a clear image of an obvious and distinctive monument, otherwise it isn't particularly accurate
It's not very accurate, but it seems consistent.
However it quite often tells me 'this is in X because the language on the sign' when there are no signs at all. Or just now I got 'The house in the background is made of wood, which is a common building material in Finland.' with a photo of a lake. There is no house, there are trees though :)
This correctly identifies South Korean landmarks, like Diamond Bridge in Busan. Since I don't have encyclopedic knowledge of world landmarks -- I wouldn't be able to recognize Diamond Bridge-like landmarks in United States -- and nearly no one does either, that alone is quite useful.
My backyard: Germany because there are trees and a fence (? Also, no).
A picture of the farmer's market of my town: correctly assume France but confidently incorrect on the town and landmarks (off by 200-300km I'd say).
Hmm interesting; one hit, one probably in vaguely the right area; both from scans of ~40 year old photos.
(As someone else noted, site is rather brokne on Firefox/Linux but does work).
I put an image of villa isnard, cascina, Pisa, and it was recognized as being from France. It is a villa with French architecture and olive trees. I then tried to upload an image still from Pisa with a building in venetian Gothic style and it was recognized being in Venice. It can be deceived quite easily imho, it looks like it just search for the corresponding architecture (maybe?) and details surrounding it but it doesn't search online. Villa isnard is quite famous (at least, you have results online) and a Google lens search would have found it
Yeah, I would guess it's identifying elements in the photo, qualifying the likelihood of those combined elements in a particular location, then outputting the assumption. I posted a picture of a Texas lake seen from a privatr residence, and it correctly guessed Texas, but pushed it off by a couple hundred miles into an Austin golf course.
Assuming this is more of a proof of concept/prototype, that's not bad. It didn't get it[1] right, but at it's core, the guess is not terrible, shift it 560km[2] south-east and you'd be bang on. I'll admit, I did set the bar a bit high.
Seems to be about as accurate as a good geoguessr player on a time limit. Recognizable vistas are generally right down to the city, and even if there's only general architecture to go off it's often right to within a couple hundred kilometers.
The explanations are a bit hit and miss. Some are great and correctly describe the names of buildings in the picture, some are only vaguely related to the picture.
Ethically this is very questionable. Of course with enough dedication humans can do the same (e.g. Rainbolt has made a Youtube career out of this), but commoditizing this for every stalker around the world has some troubling implications.
Ethics is absent in the minds of the people building and financing this. AI is about wholesale value extraction and destruction of competition done by the ecosystem of small startups repackaging AI APIs. Those APIs will be turned off once ad revenue starts flowing into the bank accounts of AI API providers.
This is close to an actual need I've managed to create for myself.
I do photography and I store those I want to share on nextcloud. In my selection and export process all metadata etc is stripped. But I realized too late that it also stripped out the geo-coordinates. No problem adding that in, but still have a laaarge amount of photos without geolocation data.
I'm too lazy to re-export all the older ones, so being able to run something like this on them would be perfect. I would be satisfied with a general area, roughly hitting the province/state its taken in. It doesn't have to be accurate at all, it's more for my own geo grouping.
This site though goes bananas on firefox/mac. Flickering and font adjustments..
> I'm too lazy to re-export all the older ones, so being able to run something like this on them would be perfect. I would be satisfied with a general area, roughly hitting the province/state its taken in. It doesn't have to be accurate at all, it's more for my own geo grouping.
I don't think this is even close to being accurate to be used in this way, out of ~10 images I uploaded it got one "correct" (right country, wrong city). Unless you want all your images to geo-tagged "Somewhere, US", probably better to re-export/re-import with your original metadata.
That's fair. I couldn't even get this to work, so not in particular looking at this implementation. I just literally was thinking of if this would be viable or not as an approach, so it was fun to see something that tries to match the bill!
If you still have the original photos, maybe you can write a script to run both the originals and exports against a perceptual hash (so as to easily identify the correct original) and then just update the JPEG EXIF data of the exports?
Depending on specific formats, you should be able to read and edit metadata without having to reprocess the images. If the exports are named similarly to the originals, you don't even need to hash them.
Funny enough it was accurate all while citing items that were not in the picture (not even cropped out), like tall buildings and signs in a specific language. I'm sure there will be refined versions that are scarily accurate. Another OSINT tool for better or worse.
FirebaseError: Installations: Create Installation request failed with error "400 INVALID_ARGUMENT: API key expired. Please renew the API key." (installations/request-failed).
I put in a photo of a small lake near Truckee, CA, several miles from Lake Tahoe, and it reported it was Lake Tahoe. It was wrong, but impressed it was geographically very close.
It's heavily biased and therefore easily tricked. I uploaded a photo from NYC.
The graffiti on the wall is a clue that the photo was taken in Detroit. The vegetation in the background is also consistent with the climate of Detroit.
It is looking for distinct features in the photo and does a probabilistic match against a tagged dataset. The features that match best on the tagged photos in its dataset are used to construct output that looks like a plausible answer. Don't use it to plan your trip.
Yeah maybe just don't do it then? If someone removes the EXIF data from a photo there's probably a reason for that, and assuming that's suspicious in some way is pretty ridiculous in a society that's supposedly all about personal freedom and the right to a fair trial.
I don't mean to be aggressive here but this seems like yet another tool that will be abused to shit by already powerful people to do even sketchier things.
I agree, this project probably shouldn't exist. But oh, well here we are and stuff like this can be built with reasonable effort. Scrape google streetview and every exif tagged Image you can get your hands on and get training.
I have no idea where this is heading, but we aren't turning back.
I'm blown away; it correctly identified a photo taken inside my house - just a picture of my kitchen - as being in eastern Massachusetts, just based on the architecture.
Maybe it does a lot better with photos related to the US, training set probably contained mostly US-related images, as only one image out of ~10 taken in various European places were correctly guessed for me. Most of the guesses was places in the US while none of the images I tried were from the US.
I gave it a photo from inside a house, you can see a person on the bed, and the white wall behind - that's it.
Obviously I wasn't expecting an accurate location, but
> This photo was taken in Los Angeles, California. We can tell this from the architecture of the buildings in the background, as well as the vegetation. The palm trees are a dead giveaway that this is Los Angeles.
There are no palm trees, the photo wasn't taken in the US and palm trees exist outside of LA.
I also fed a photo of some quite distinctive castle ruins. It mislocated that by 100s of miles.
I gave it a picture from a bar in Austin. It nailed it, but with some interesting hallucinations in the description. The photo had a small Texas flag, but nobody was wearing cowboy hats, and there was nothing with "Austin" on it in the photo. Description was:
This photo was taken inside a bar. There are several clues that indicate this is Austin. First, there is a sign on the wall that says "Austin." Second, there is a Texas flag on the wall. Third, there are several people wearing cowboy hats, which is a common sight in Austin. The coordinates of this photo are
105 comments
[ 2.3 ms ] story [ 171 ms ] threadI did manage to get it to place a picture of a praying mantis I took in Japan to be from California...
See this thread: https://news.ycombinator.com/item?id=40233248
Don't recommend.
That was with an image I took in London on my phone.
https://i.imgur.com/ywc1Hn0.png
(Chrome on Windows)
[1] https://imgur.com/a/67t0TVt
[2] https://imgur.com/a/aspf8px
The explanations are a bit hit and miss. Some are great and correctly describe the names of buildings in the picture, some are only vaguely related to the picture.
Ethically this is very questionable. Of course with enough dedication humans can do the same (e.g. Rainbolt has made a Youtube career out of this), but commoditizing this for every stalker around the world has some troubling implications.
I do photography and I store those I want to share on nextcloud. In my selection and export process all metadata etc is stripped. But I realized too late that it also stripped out the geo-coordinates. No problem adding that in, but still have a laaarge amount of photos without geolocation data.
I'm too lazy to re-export all the older ones, so being able to run something like this on them would be perfect. I would be satisfied with a general area, roughly hitting the province/state its taken in. It doesn't have to be accurate at all, it's more for my own geo grouping.
This site though goes bananas on firefox/mac. Flickering and font adjustments..
I don't think this is even close to being accurate to be used in this way, out of ~10 images I uploaded it got one "correct" (right country, wrong city). Unless you want all your images to geo-tagged "Somewhere, US", probably better to re-export/re-import with your original metadata.
https://github.com/JohannesBuchner/imagehash
Depending on specific formats, you should be able to read and edit metadata without having to reprocess the images. If the exports are named similarly to the originals, you don't even need to hash them.
I don't mean to be aggressive here but this seems like yet another tool that will be abused to shit by already powerful people to do even sketchier things.
I have no idea where this is heading, but we aren't turning back.
Good
> Rural area
good
> [pin in the center of Stockholm, the most urban area in Sweden]
ouch.. not so good.
I gave it a photo from inside a house, you can see a person on the bed, and the white wall behind - that's it.
Obviously I wasn't expecting an accurate location, but
> This photo was taken in Los Angeles, California. We can tell this from the architecture of the buildings in the background, as well as the vegetation. The palm trees are a dead giveaway that this is Los Angeles.
There are no palm trees, the photo wasn't taken in the US and palm trees exist outside of LA.
I also fed a photo of some quite distinctive castle ruins. It mislocated that by 100s of miles.
This photo was taken inside a bar. There are several clues that indicate this is Austin. First, there is a sign on the wall that says "Austin." Second, there is a Texas flag on the wall. Third, there are several people wearing cowboy hats, which is a common sight in Austin. The coordinates of this photo are
Even photos with signs in Irish were pointing to England, it's half funny, half offensive.