Viewer: https://tubvsig-so2sat-vm1.srv.mwn.de/ (might already be hugged to death? When it works you get a heatmap when zoomed out, and 3d models with flat roofs when you zoom in far enough)
Looks pretty cool. Obviously only covers above-ground buildings or above-ground portions of buildings. Also seems like it tends to view buildings built wall-to-wall next to each other as the same building, but not always. So if you calculate something like average building volume you are bound to be off quite a bit
InfoReseaux remarked that this data is suspicious, to say the least: CC BY-NC 4.0 but contains ODbL licensed data coming from Openstreetmap - but also from Microsoft.
Christian Quest sent the following message to the authors:
"I’m writing to you because I’m surprised by the choice of data license you’ve set on the GlobalBuildingAtlas dataset.
As mentioned and explained in your paper, at least two data sources you’ve been using to create this dataset are under the Open Database License (ODbL): OpenStreetMap and Microsoft building datasets.
I’ve downloaded the extract of data you’re proposing to have a look at the final dataset, and it confirms that building polygons from OSM (and Microsoft) are present in the resulting dataset in a substantial portion.
In such case, your dataset must be published under the ODbL licence (see 4.2), because it is a derivative database (see 1.0 of ODbL license for definition).
A copy of this message has also been sent to the Legal Working Group of the OSM Foundation.
Thanks in advance to fix quickly the license of the dataset you published. This will also allow OpenStreetMap contributors to use it to improve OpenStreetMap, which is not possible with the CC-BY-NC you choose."
So here is finally the data that I needed for my idea: walk directions from A to B while never leaving the shadows, suggesting bars and pubs where to wait while the shadows catch up and let you cross “safely” where before was sunny.
Useful for scorching weather places like south of Spain.
So running directions from B to A while never touching any shade, suggesting churches and rehab centers to ignore while the shadows stay perfectly still?
Currently building something like this for my city. One major challenge I've come across so far is that most APIs will give you coordinates for an address of a place within a building but none of the free, paid or crowdsourced options have reliable information on outside seating polygons. Of course, you could always display the places around you with live shadow data on a map, leaving it up to the user to zoom in and decide based on the satellite image whether the restaurant or café offers outside seating. But to plot the route and then suggest nearby sun/shadow seating options to the user, you would need this information.
You'd need stereo fotos from a low flying plane for this. Several cities are doing this for decades already. I've built such a 3D map for my city like 25 years ago.
Are you referring to the outside seating polygons? Wouldn’t these stereo images still have a lot of noise (trees, cars, trucks, smaller non-building structures) obstructing the target areas?
The height measure when you click on a building is not very accurate. Clicked on two next to each other (one single floor and one with two floors) and the single story house was marked taller by the site.
Can confirm. My garage is marked double the height of my actual house. Both are incorrect. My shop building on my property, next to the garage is actually missing.
The positioning of the shapes in relation to each other are also wrong. And not in a subtle way.
Looked at the house my father built in 1985. A few hundred kilometers from my current location. The shape is wrong (as are the shapes of all neighboring buildings. As are the positions toward each other (distance between houses, rotation of shapes). The heights are also significantly wrong. The two story houses on the opposite side of the road are said to be slightly above 2 meters in height.
I've had a project idea in the back of my mind for a long time: take all of my locations that I've saved in Swarm (https://swarmapp.com) and 3D render all of the buildings that I've been to. The issue was a lack of models of the actual buildings - which is here! Maybe it is possible...
They've used a vision transformer to estimate building heights from monocular aerial photographs, so they're guesses at best. Calling this a map is a stretch.
What is the crowd process to upgrade/correct this? I notice massive errors in small Midwest US towns (multiple buildings being combined/ommitted, etc)?
Looking around my home area, I found HS football bleachers showing up as buildings and an outdoor basketball court showing up as a building. Looking at the coloring of those features on Google maps, I can see how the mistakes might have been made. But still feels like it needs tuning.
To be fair, I checked my boyhood home in very rural Maine and it was correct for the size and shape of the multiple farm buildings.
Yeah some areas I looked at seemed fine (US, EU, Shangahi), but some the buildings are on top of the roads, and they look more like Perlin Noise than actual buildings. E.g main parts of Beijing.
OpenStreetMap and such have pretty good building footprints, sometimes enriched with building heights. Maybe that can be used for some correction
Looking at where I live and where I grew up the building heights are quite badly estimated.
- Some groups of houses around here that are more or less identically built but on sloped terrain are reported to have widly differing heights
- My neighbour building is reported to be half the height of this building (they're more or less equally high at 5 stories)
- A small office shack behind the neighbour building is reported to be taller than it (it's a single-story building, the neighbour building is 5 stories)
- The freestanding buildings on the farm where I grew up are like you said, badly combined, much of the estimation there seems to be dependent on shadows,etc.
Or, they subtracted a digital elevation model from a digital surface model, ran a point-in-polygon match against an existing building dataset, and labelled the difference as the height of the building. No ML needed.
Still useful for a sense of building density. If almost everything called out is some sort of construction, then the density map of the world is a realistic estimate of human occupation.
You have uBlock Origin or another ad blocker installed, their cookie banner disables scrolling but your ad blocker is blocking the cookie banner (same thing happened to me)
Same here with uBlock on iOS. This happens every now and then, not too often. But when it does, I usually decide the page is probably not worth reading anyway.
I'd draw the line at boats that are primarily used when docked. A house boat, a restaurant located on a permanently docked boat or Russia's floating power plants are all buildings in my book
Alphabet, please open your mapping data. I understand it’s your IP, but given that the value in Google Maps far exceeds the mapping data and won’t be replaced, think of all of the good that would come of it. For one thing, you wouldn’t have to worry about people posting inaccurate maps like these.
For sure not all buildings. My first query was "Costa del Este, Juan Díaz, Distrito de Panamá, Panamá Province, Panama" and no coverage. This region is packed with high rises.
Yup, check my home neighborhood and it is good coverage, but checked a rural area I know has houses around a lake (not just tiny fishing shacks, full-sized 200+m^2 houses) with addresses and electrical and telecom/internet service, but there there are no buildings shown. Plus, even though there is a lot of tree coverage, many of the houses are clearly visible in Google Satellite View, so it is unclear why they wouldn't have been picked up.
The reason, I would guess, is that it takes at least a couple satellite passes to be able to get the 3D data, and then a processing step to actually obtain it, and they've not done that for sparsely populated areas.
Revealing: third-world areas look different. Curiously, Australia resembles third-world building density more than it resembles the industrialized world.
If it was able to detect roughly how many buildings there are in that area, that’s already fairly useful at global scale. It seems they are focused on measuring urbanization. It’s good to know though that it may be dodgy at that level of granularity.
Indeed, my neighborhood doesn't show any buildings at all even though I know all of our houses have been there at least 30 years and were still there when I left this morning.
I haven't visited the map, but I wouldn't be surprised to learn they didn't attempt to re-create the building shapes, just location and estimated size. The real shape seems pointless for this dataset.
Building shapes are actually pretty accurate in urban areas. I haven't read the paper, but I think what they were after is building volume. Which obviously requires the footprint shape and height
Github: https://github.com/zhu-xlab/GlobalBuildingAtlas
Viewer: https://tubvsig-so2sat-vm1.srv.mwn.de/ (might already be hugged to death? When it works you get a heatmap when zoomed out, and 3d models with flat roofs when you zoom in far enough)
Looks pretty cool. Obviously only covers above-ground buildings or above-ground portions of buildings. Also seems like it tends to view buildings built wall-to-wall next to each other as the same building, but not always. So if you calculate something like average building volume you are bound to be off quite a bit
InfoReseaux remarked that this data is suspicious, to say the least: CC BY-NC 4.0 but contains ODbL licensed data coming from Openstreetmap - but also from Microsoft.
Thread on the Openstreetmap forum: https://community.openstreetmap.org/t/is-globalbuildingatlas...
Christian Quest sent the following message to the authors:
"I’m writing to you because I’m surprised by the choice of data license you’ve set on the GlobalBuildingAtlas dataset.
As mentioned and explained in your paper, at least two data sources you’ve been using to create this dataset are under the Open Database License (ODbL): OpenStreetMap and Microsoft building datasets.
I’ve downloaded the extract of data you’re proposing to have a look at the final dataset, and it confirms that building polygons from OSM (and Microsoft) are present in the resulting dataset in a substantial portion.
In such case, your dataset must be published under the ODbL licence (see 4.2), because it is a derivative database (see 1.0 of ODbL license for definition).
A copy of this message has also been sent to the Legal Working Group of the OSM Foundation.
Thanks in advance to fix quickly the license of the dataset you published. This will also allow OpenStreetMap contributors to use it to improve OpenStreetMap, which is not possible with the CC-BY-NC you choose."
Yes, this dataset is not CC BY-NC 4.0. Be careful.
So here is finally the data that I needed for my idea: walk directions from A to B while never leaving the shadows, suggesting bars and pubs where to wait while the shadows catch up and let you cross “safely” where before was sunny.
Useful for scorching weather places like south of Spain.
I do a similar game when walking around but the complete inverse. I live in the UK.
So running directions from B to A while never touching any shade, suggesting churches and rehab centers to ignore while the shadows stay perfectly still?
You can fool me Dracula
Currently building something like this for my city. One major challenge I've come across so far is that most APIs will give you coordinates for an address of a place within a building but none of the free, paid or crowdsourced options have reliable information on outside seating polygons. Of course, you could always display the places around you with live shadow data on a map, leaving it up to the user to zoom in and decide based on the satellite image whether the restaurant or café offers outside seating. But to plot the route and then suggest nearby sun/shadow seating options to the user, you would need this information.
You'd need stereo fotos from a low flying plane for this. Several cities are doing this for decades already. I've built such a 3D map for my city like 25 years ago.
Are you referring to the outside seating polygons? Wouldn’t these stereo images still have a lot of noise (trees, cars, trucks, smaller non-building structures) obstructing the target areas?
Outer walls, but mostly roof lines.
Some noise, but since houses have a certain size and being mostly rectangular helps. Just carports and garden houses are special.
Would need this, inverted, for winter.
Check out sunseekr its popular here in the uk: https://sunseekr.com/
In that case https://app.shadowmap.org could be a better recommendation.
The height measure when you click on a building is not very accurate. Clicked on two next to each other (one single floor and one with two floors) and the single story house was marked taller by the site.
Can confirm. My garage is marked double the height of my actual house. Both are incorrect. My shop building on my property, next to the garage is actually missing.
The positioning of the shapes in relation to each other are also wrong. And not in a subtle way.
Looked at the house my father built in 1985. A few hundred kilometers from my current location. The shape is wrong (as are the shapes of all neighboring buildings. As are the positions toward each other (distance between houses, rotation of shapes). The heights are also significantly wrong. The two story houses on the opposite side of the road are said to be slightly above 2 meters in height.
quite a lot of height errors, checked my neighborhood as well, often by orders of magnitude off.
> This is a huge leap from the previous global dataset, which contained about 1.7 billion buildings
I use the monthly release of Overture Maps building dataset and the last one had around 2.3 billion buildings iirc
[1] https://docs.overturemaps.org/blog/2025/11/19/release-notes/
I've had a project idea in the back of my mind for a long time: take all of my locations that I've saved in Swarm (https://swarmapp.com) and 3D render all of the buildings that I've been to. The issue was a lack of models of the actual buildings - which is here! Maybe it is possible...
Google recently has made efforts to link Gemini to Google Earth data. “Earth AI”
https://ai.google/earth-ai/
Has anyone tried this? I’m not quite sure how to prompt it effectively.
arxiv:2510.18318
They've used a vision transformer to estimate building heights from monocular aerial photographs, so they're guesses at best. Calling this a map is a stretch.
I checked the building heights where I live vs the actual heights, sometimes they display the height as half of what the actual height is.
I found a few that are off by a factor of about 4. Always lower than in reality, never taller.
Hey maps used to have sea serpents and winged griffins. It's a map. Maybe not the best map.
What is the crowd process to upgrade/correct this? I notice massive errors in small Midwest US towns (multiple buildings being combined/ommitted, etc)?
Looking around my home area, I found HS football bleachers showing up as buildings and an outdoor basketball court showing up as a building. Looking at the coloring of those features on Google maps, I can see how the mistakes might have been made. But still feels like it needs tuning.
To be fair, I checked my boyhood home in very rural Maine and it was correct for the size and shape of the multiple farm buildings.
Yeah some areas I looked at seemed fine (US, EU, Shangahi), but some the buildings are on top of the roads, and they look more like Perlin Noise than actual buildings. E.g main parts of Beijing.
OpenStreetMap and such have pretty good building footprints, sometimes enriched with building heights. Maybe that can be used for some correction
It's machine-learning generated "slop" honestly.
Looking at where I live and where I grew up the building heights are quite badly estimated.
- Some groups of houses around here that are more or less identically built but on sloped terrain are reported to have widly differing heights
- My neighbour building is reported to be half the height of this building (they're more or less equally high at 5 stories)
- A small office shack behind the neighbour building is reported to be taller than it (it's a single-story building, the neighbour building is 5 stories)
- The freestanding buildings on the farm where I grew up are like you said, badly combined, much of the estimation there seems to be dependent on shadows,etc.
Or, they subtracted a digital elevation model from a digital surface model, ran a point-in-polygon match against an existing building dataset, and labelled the difference as the height of the building. No ML needed.
There's a notice in the bottom-left corner on desktop that says: "This is a machine-learning-derived product. Errors may occur"
Other comments said they fed 2d aerial imagery into transformer and thats it.
Still useful for a sense of building density. If almost everything called out is some sort of construction, then the density map of the world is a realistic estimate of human occupation.
At least in Firefox, this page doesn't scroll at all. No scrollbar, scroll wheel/cursor keys/PageDown all do nothing.
How does someone screw up CSS so completely?
You have uBlock Origin or another ad blocker installed, their cookie banner disables scrolling but your ad blocker is blocking the cookie banner (same thing happened to me)
Same here with uBlock on iOS. This happens every now and then, not too often. But when it does, I usually decide the page is probably not worth reading anyway.
For me it was the cookie notice. Had to disable uBlock and reject them before the page could be responsive again.
works for me on FF
Groups of boats in the dock have become buildings. Cool project still.
If AI says a floating building is a building I will agree with it. It isn't wrong but it isn't right.
I'd draw the line at boats that are primarily used when docked. A house boat, a restaurant located on a permanently docked boat or Russia's floating power plants are all buildings in my book
Can’t wait for the "falsehoods programmers believe about buildings" blog post
lol, AI lacking effective QA process it seems
Alphabet, please open your mapping data. I understand it’s your IP, but given that the value in Google Maps far exceeds the mapping data and won’t be replaced, think of all of the good that would come of it. For one thing, you wouldn’t have to worry about people posting inaccurate maps like these.
For sure not all buildings. My first query was "Costa del Este, Juan Díaz, Distrito de Panamá, Panamá Province, Panama" and no coverage. This region is packed with high rises.
Yup, check my home neighborhood and it is good coverage, but checked a rural area I know has houses around a lake (not just tiny fishing shacks, full-sized 200+m^2 houses) with addresses and electrical and telecom/internet service, but there there are no buildings shown. Plus, even though there is a lot of tree coverage, many of the houses are clearly visible in Google Satellite View, so it is unclear why they wouldn't have been picked up.
Cool stuff, but they still have work to do.
The reason, I would guess, is that it takes at least a couple satellite passes to be able to get the 3D data, and then a processing step to actually obtain it, and they've not done that for sparsely populated areas.
Where?
Revealing: third-world areas look different. Curiously, Australia resembles third-world building density more than it resembles the industrialized world.
I wonder when people will be able to make a map of all the people in the world at a particular moment. Pretty soon I think.
All of the houses on my block are very weird shapes on this map, it doesn't reflect reality very well
If it was able to detect roughly how many buildings there are in that area, that’s already fairly useful at global scale. It seems they are focused on measuring urbanization. It’s good to know though that it may be dodgy at that level of granularity.
Indeed, my neighborhood doesn't show any buildings at all even though I know all of our houses have been there at least 30 years and were still there when I left this morning.
When you left this morning, sure, but who knows what’s happened in the interim.
I haven't visited the map, but I wouldn't be surprised to learn they didn't attempt to re-create the building shapes, just location and estimated size. The real shape seems pointless for this dataset.
Building shapes are actually pretty accurate in urban areas. I haven't read the paper, but I think what they were after is building volume. Which obviously requires the footprint shape and height
Looks like it has trouble depicting the height of the highest buildings like Burj Khalifa and Merdeka 118
Would be great to somehow incorporate this into a game where you can play in any city in the world.
The heights seem all over the place. Look at the WTC, for example.
It seemed quite accurate for my region (São Paulo downtown).
oh wow basically amazed
[flagged]
How do you figure this? Russia's buildings aren't being demolished.
Many are abandoned, many lack the even most basic features, like plumbing. Hell, there is not even paved roads once you leave the bigger cities.
Some are. There is a war on.