This is pretty cool, but I feel as a pokehunter (Pokemon Go player), I have been tricked into working to contribute training data so that they can profit off my labor. How? They consistently incentivize you to scan pokestops (physical locations) through "research tasks" and give you some useful items as rewards. The effort is usually much more significant than what you get in return, so I have stopped doing it. It's not very convenient to take a video around the object or location in question. If they release the model and weights, though, I will feel I contributed to the greater good.
This title is editorialized. The real title is: "Building a Large Geospatial Model to Achieve Spatial Intelligence"
> Otherwise please use the original title, unless it is misleading or linkbait; don't editorialize.
My personal layman's opinion:
I'm mostly surprised that they were able to do this. When I played PokƩmon GO a few years back, the AR was so slow that I rarely used it. Apparently it's so popular and common, it can be used to train an LGM?
I also feel like this is a win-win-win situation here, economically. Players get a free(mium) game, Niantic gets a profit, the rest of the world gets a cool new technology that is able to turn "AR glasses location markers" into reality. That's awesome.
Not wanting to over-do it, but is there possibly an argument the data about geospatial should be in the commons and google have some obligation to put the data back into the commons?
I'm not arguing to a legal basis but if it's crowdsourced, then the inputs came from ordinary people. Sure, they signed to T&Cs.
Philosophically, I think knowledge, facts of the world as it is, even the constructed world, should be public knowledge not an asset class in itself.
I can really imagine a meeting with the big brasses of Google/Niantic a few years ago that went along
- We need to be the first to have a better, new generation 3D model of the world to build the future of maps on it. How can we get that data?"
+ What about gamifying it and crowd-sourcing it to the masses?
- Sure! Let's buy some Pokemon rights!
It's scary but some people do really have some long-term vision
I'm sure the CIA already has access. [1] People were raising privacy concerns years ago. [2]
[1] https://www.networkworld.com/article/953621/the-cia-nsa-and-...
[2] https://kotaku.com/the-creators-of-pokemon-go-mapped-the-wor...
Brian Maclendon (Niantic) presented some interesting details about this in his recent Bellingfest presentation:
https://www.youtube.com/live/0ZKl70Ka5sg?feature=shared&t=12...
> For example, it takes us relatively little effort to back-track our way through the winding streets of a European old town. We identify all the right junctions although we had only seen them once and from the opposing direction.
That is true for some people, but I'm fairly sure that the majority of people would not agree that it comes naturally to them.
Impressive, but this is one of those "if this is public knowledge, how far ahead is the _not_ public knowledge" things
This is literally what I built my first company around starting in 2012, when Niantic was still working on Ingress
I describe it here during 500 Startups demo day: https://youtu.be/3oYHxdL93zE?si=cvLob-NHNEIJqYrI&t=6411
I further described it on the Planet of the Apps episode 1
Here's my patent from 2018: https://patents.google.com/patent/US10977818B2/en
So. I'm not really sure what to do here given that this was exactly and specifically what we were building and frankly had a lot of success in actually building.
Quite frustrating
Interestingly, Pokemon GO only prompts players to scan a subset of the Points of Interest on the game map. Players can manually choose to scan any POI, but with no incentive for those scans I'm sure it almost never happens.
> Today we have 10 million scanned locations around the world, and over 1 million of those are activated and available for use with our VPS service.
This 1 in 10 figure is about accurate, both from experience as a player and from perusing the mentioned Visual Positioning System service. Most POI never get enough scan data to 'activate'. The data from POI that are able to activate can be accessed with a free account on Niantic Lightship [1], and has been available for a while.
I'll be curious to see how Niantic plans to fill in the gaps, and gather scan data for the 9 out of 10 POI that aren't designated for scan rewards.
Somehow I always thought something like that would have been the ultimate use case for Microsoft Photosynth (developed from Photo Tourism research project), ideally with a time dimension, like browsing photos in a geo spatio-temporal context.
I expect that was also some reason behind their flickr bid back then.
https://medium.com/@dddexperiments/why-i-preserved-photosynt...
https://phototour.cs.washington.edu
https://en.wikipedia.org/wiki/Photosynth
at least any patents regarding this will also expire about 2026.
I still don't get what LGM is. From what I understood, it isn't actually about any "geospatial" data at all, is it? It is rather about improving some vision models to predict how the backside of a building looks, right? And training data isn't of people walking, but from images they've produced while catching pokemons or something?
P.S.: Also, if that's indeed what they mean, I wonder why having google street view data isn't enough for that.
Iāve published research in this general arena and the sheer amount of data they need to get good is massive. They have a moat the size of an ocean until most people have cameras and depth sensors on their face
Itās funny, we actually started by having people play games as well but we expressly told them it was to collect data. Brilliant to use an AR game that people actually play for fun
I'm guessing this can be the new bot that could play competitively at GeoGuesser. It would be interesting if Google trained a similar model and released it using all the Street Map data, I sure hope they do.
Has anyone done something similar with the geolocated WIFI MAC addresses, to have small model for predicting location from those.
> Today we have 10 million scanned locations around the world, and over 1 million of those are activated and available for use with our VPS service. We receive about 1 million fresh scans each week
Wait, they get a million a week but they only have a total of 10 million, ie 10 days worth? Is this a typo or am I missing something?
I wonder how this can be combined with satellite data, if at all?
Genuinely impressed Google had the vision and resources to commit to a 10 year data collection project
So what they are doing is not different from previous "VPS" systems, its how they are doing it.
What is a "VPS" At its heart, Visual Positioning Systems are actually pretty simple. You build a 3d point cloud of a place, with each point being a repeatable unique feature that can be extracted from an image (see https://blog.ekbana.com/extracting-invariant-features-from-i...) Basically a "finger print"/landmark of a thing in real life that can be extracted from an image reliably.
To make that work, you need to generate a large map of these points: https://www.researchgate.net/figure/Sparse-point-cloud-Figur... Which basically involves taking lots of pictures with GPS tags on where they are. Google has the advantage of street view, Niantic has it's game. Others had to pay a bunch of people to go round a city with cameras.
Once you build that pointcloud (which isn't actually that easy, you can't do it all at once, and aligning point clouds is hard.) you can then use trigonometry to work out where a picture is. This is called "re-localization" which is a stupid name. The hard part is the data management. There are billions of points in the world, partitioning the database so that you can quickly locate a picture is the hard part.
Hence this approach, which is basically "train a model to do it for us" You still get a "VPS", you still need all that data, but they hope that a model will able to optimize for speed.
is it private?
No, the original system isn't private. If they've done their job properly, then nothing identifiable will be in the "map" as thats extra data you dont need. What they do with the raw photos, and the metadata that they contain is another matter.
Even before LLMs, I knew they are going to launch a fine grained mapping service with all that camera and POI data. Now this one is actually much better obviously. Very few companies actually have this kind of data. Remains to be seen how they make money out of this
Very cool.
However, I can't fully agree that generating 3d scene "on the fly" is the future of maps and many other use cases for AR.
The thing with geospatial, buildings, roads, signs, etc. objects is that they are very static, not many changes are being made to them and many changes are not relevant to the majority of use cases. For example: today your house is white and in 3 years it has stains and yellowish color due to time, but everything else is the same.
Given that storage is cheap and getting cheaper, bandwidth of 5G and local networks is getting too fast for most current use cases, while computer graphics compute is still bound by our GPU performance, I say that it would be much more useful to identify the location and the building that you are looking at and pull the accurate model from the cloud (further optimisations might be needed like to pull only the data user has access to or needs access to given the task he is doing). Most importantly users will need to have access to a small subset of 3D space on daily basis, so you can have a local cache on end devices for best performance and rendering. Or stream rendered result from the cloud like nVidia GDN is doing.
Most precise models will come from CAD files for newly built buildings, retrospectively going back to CAD files of buildings build in last 20-30 years(I would bet most of them have some soft of computer model made before) and finally going back even further - making AI look at the old 2D construction plans of the building and reconstructing it in 3D.
Once the building is reconstructed (or a concrete pole like shown in the article) you can pull its 3D model from the cloud and place it in front of the user - this will cover 95% of use cases for AR. For 5% of the tasks you might want real time recognition of the current state of surfaces for some tasks or changes in geometry (like tracking the changes in the road quality compared with the previous scans or with reference model), but these cases can be tackled separately and having precise 3D model will only help, but won't be needed to be reconstructed from scratch.
This is a good 1st step to make a 3D map, however there should be an option to go to the real location and make edits to 3D plan by the expert so that the model can be precise and not "kind of" precise.
Applications that I thought of as I read this:
Real-Time mapping of the environment for VR experiences with built-in semantic understanding.
Winning at geoguesser, automated doxing of anybody posting a picture of themselves.
Robotic positioning and navigation
Asset generation for video games. Think about generating an alternate New York City that's more influenced by Nepal.
I'm getting echoes of neural radiance fields as well.
Procedural generation of an alternative planet is the kind of stuff that the No Man's sky devs could only dream of.
So that's why Pokemon was notoriously impactful on battery life. They were recording and uploading our videos the whole time?
I wonder if there's a sweet spot for geospatial model size.
A model trained on all data for 1m in every direction would probably be too sparse to be useful, but perhaps involving data from a different continent is costly overkill? I expect most users are only going to care about their immediate surroundings. Seems like an opportunity for optimization.
Waymo is supposedly geofenced because they need detailed maps of an area. And this is supposedly a blocker for them deploying everywhere. But then Google goes and does something like this, and I'm not sure, if it's even really true that Waymo needs really detailed maps, that it's an insurmountable problem.
Conversation about āplayers are the productā of PokĆ©mon go asideā¦ What are some practical applications of an LGM?
Seems like navigation is āsolvedā? Thereās already a lot of technology supporting permanence of virtual objects based on spatial mapping? Better AI generated animations?
I am sure there are a ton of innovations it could unlockā¦
It may not be Geospatial data at all and I'm not sure how much the users consented but the data collection strategy was well crafted. I remember recommending building a game to collect handwriting data from testers (about a thousand), to the research lab I worked for long time back.
This looks like another use of data not following the original purpose of the collected data. Clearly it should be illegal to use any such data without asking every single user whose data they want to use for consent. And by that I do not mean some extortion scheme.
It seems I was unable to generate the image for the "SWAT ESPORT" logo at this time. Let me know if you would like me to try again or if you'd like to adjust the description.
Is this related to NeRF (neural radiance fields)?
This seems like itād be quite handy to have in an autonomous vehicle of any kind
Iām intrigued by the generative possibilities of such a model even more than how it could be used with irl locations. Imagine a game or simulation that creates a realistic looking American suburbia on the fly. It honestly canāt be that difficult, it practically predicts itself.
The data marginally better than what google already have
Don't quite understand the application of this?
People complaining here that you are somehow owed something for contributing to the data set, or that because you use google maps or reCAPTCHA you are owed access to their training data. I mean, I'd like that data too. But you did get something in return already. A game that you enjoy (or your wouldn't play it), free and efficient navigation (better than your TomTom ever worked), sites not overwhelmed by bots or spammers. Yeah google gets more out of it than you probably do, but it's incorrect to say that you are getting 'nothing' in return.
Lunduke is happy: āI told you so!ā
Going to try to clear this up from speculation as best I can.
Niantic was a spinoff divested from Google Maps roughly a decade ago who created a game called Ingress. This used Open Street Maps data to place players in the real world and they could designate locations as points of interest (POI), which Niantic used human moderators to judge as sufficiently noteworthy. Two years after Ingress was released, Niantic purchased limited rights to use Pokemon IP and bootstrapped Pokemon Go from this POI data. Individual points of interest became Pokestops and Gyms. Players had to physically go to these locations and they could receive in-game items needed to continue playing or battle other Pokemon.
From the beginning, Pokemon Go had AR support, but it was gimmicky and not widely used. Players would post photos of the real world with Pokemon overlaid and then turn it off, as it was a significant battery drain and only slowed down your ability to farm in-game items. The game itself has always been a grind type of game. Play as much as possible to catch Pokemon, spin Pokestops, and you get rewards from doing so. Eventually, Niantic started having raids as the only way to catch legendary Pokemon. These were multiplayer in-person events that happened at prescribed times. A timer starts in the game and players have to be at the same place at the same time to play together to battle a legendary Pokemon, and if they defeat it, they'll be rewarded with a chance to catch one.
Something like a year after raids were released, Niantic released research tasks as a way to catch mythical Pokemon. These required you to complete various in-game tasks, including visiting specific places. Much later than this, these research tasks started to include visiting designated Pokestops and taking video footage, from a large enough variety of angles to satisfy the game, and then uploading that. They started doing this something like four or five years ago, and getting any usable data out of it must have required an enormous amount of human curation, which was largely volunteer effort from players themselves who moderated the uploads. The game itself would give you credit simply for having the camera on while moving around enough, and it was fairly popular to simply videotape the sidewalk and the running game had no way to tell this was not really footage of the POI.
The quality of this data has always been limited. Saying they've managed to build local models of about 1 million individual objects leaves me wondering what the rate of success is. They've had hundreds of millions of players scanning presumably hundreds of millions of POI for half a decade. But a lot of the POI no longer exist. Many of them didn't exist even when Pokemon Go was released. Players are incentivized to have as many POI near them as possible because this provides the only way to actually play, and Niantic is incentivized to leave as much as they can in the game and continually add more POI because, otherwise, nobody will play. The mechanics of the game have always made it tremendously imbalanced in that living near the center of a large city with many qualifying locations results in rich, rewarding gameplay, whereas living out in the suburbs or a rural area means you have little to do and no hope of ever gaining the points that city players can get.
This means many scans are of objects that aren't there. Near me, this includes murals that have long been painted over, monuments to confederate heroes that were removed during Black Lives Matter furors of recent years, small pieces of art like metal sculptures and a mailbox decorated to look like Spongebob that simply are not there any more for one reason or another, but the POI persist in the database anyway. Live scans will show something very different from the original photo that still shows up in-game to tell you what the POI is.
Another problem is many POI can't be scanned from all sides. They're behind fences, closed off because of construction, or otherwise obstructed.
Yet another problem is GPS drift. I live near downtown Dallas right now, but when the game started, I lived smack dab in the city center, across the street from AT&T headquarters. I started playing as something to do when walking during rehab from spine surgeries, but I was often bedridden and couldn't actually leave the apartment. No problem. I could receive sometimes upwards of 50km a day of credit for walking simply by leaving my phone turned on with the game open. As satellite line of sight is continually obstructed and then unobstructed by all the tall buildings surrounding your actual location, your position on the map will jump around. The game has a built-in speed limit meant to prevent people from playing while driving, and if you jump too fast, you won't get credit, but as long as the jumps in location are small enough to keep your average over some sampling interval below that limit, you're good to go. Positions within a city center where most of the POI actually are is very poor.
They claim here that they have images from "all times of day," which is possibly true if they literally mean daylight hours. I'm awake here writing this comment at 2:30 AM and have always been a very early riser. I stopped playing this game last summer, but when I still played, it was mostly in darkness, and one of the reason I quit was the frustration of constantly being given research tasks I could not possibly complete because the game would reject scans made in the dark.
Finally, POI in Ingress and Pokemon Go are all man-made objects. Whatever they're able to get out of this would be trained on nothing from the natural world.
Ultimately, I'm interested in how many POI the entire map actually has globally and what proportion the 1 million they've managed to build working local models of represents. Seemingly, it has to be objects that (1) still exist, (2) are sufficiently unobstructed from all sides, and (3) in a place free from GPS obstructions such that the location of players on the map is itself accurate.
That isn't nothing, but I'm enormously skeptical that they can use this to build what they're promising here, a fully generalizable model that a robot could use to navigate arbitrary locations globally, as opposed to something that can navigate fairly flat city peripheries and suburbs during daylight hours. If Meta can really get a large enough number of people to wear sunglasses with always-on cameras on them, this kind of data will eventually exist, but I highly doubt what Niantic has right now is enough.
We do this at [name redacted as this was a joke].
When users scan their barcode, the preview window is zoomed in so users think its mostly barcode. We actually get quite a bit more background noise typically of a fridge, supermarket aisle, pantry etc. but it is sent across to us, stored, and trained on.
Within the next year we will have a pretty good idea of the average pantry, fridge, supermarket aisle. Who knows what is next
[dead]
Fucking cool. Hi old Niantic teammates ;).
The cia has to be all over this.
I'm confused by both this blog post, and the reception on HN. They... didn't actually train the model. This is an announcement of a plan! They don't actually know if it'll even work. They announced that they "trained over 50 million neural networks," but not that they've trained this neural network: the other networks appear to just have been things they were doing anyway (i.e. the "Virtual Positioning Systems"). They tout huge parameter counts ("over 150 trillion"), but that appears to be the sum of the parameters of the 50 million models they've previously trained, which implies each model had an average of... 3MM parameters. Not exactly groundbreaking scale. You could train one a single consumer GPU.
This is a vision document, presumably intended to position Niantic as an AI company (and thus worthy of being showered with funding), instead of a mobile gaming company, mainly on the merit of the data they've collected rather than their prowess at training large models.