How could field collection for OSM look like in the future?
This is a modal window.
The media could not be loaded, either because the server or network failed or because the format is not supported.
Formal Metadata
| Title |
| |
| Title of Series | ||
| Number of Parts | 41 | |
| Author | ||
| License | CC Attribution 3.0 Unported: You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor. | |
| Identifiers | 10.5446/58206 (DOI) | |
| Publisher | ||
| Release Date | ||
| Language |
Content Metadata
| Subject Area | ||
| Genre | ||
| Abstract |
|
State of the Map US 202232 / 41
1
4
7
8
14
18
21
22
24
27
28
30
37
39
00:00
BuildingDisk read-and-write headMappingMultiplication signScaling (geometry)GeometryHand fanMedical imagingSpreadsheetScripting languageLevel (video gaming)Conditional-access moduleGoodness of fitRight angleAddress spaceDivision (mathematics)Service (economics)Uniform resource locatorBoss CorporationMassPosition operatorPredictabilityProcess (computing)Image resolutionMobile appClosed setView (database)MultilaterationEndliche ModelltheorieContent (media)Limit (category theory)Inflection pointValue-added networkTransportation theory (mathematics)Device driverDatabase normalizationWave packetInheritance (object-oriented programming)Operator (mathematics)Google Street ViewSpacetimeCodeTask (computing)Motion blurComputer programmingCASE <Informatik>Machine visionoutputOrder (biology)NavigationNeuroinformatikGame controllerConfiguration spaceBitComplete metric spacePower (physics)Core dumpPlastikkarteElectronic program guideNumberParallel portArithmetic meanPoint cloudLaptopState of matterRow (database)TrailPoint (geometry)QuicksortData managementData storage deviceSoftwareInformation privacyMereologyIterationCuboidComputer hardwareRevision controlField (computer science)Sound effectMusical ensembleProfil (magazine)CoprocessorDialectMeta elementDifferent (Kate Ryan album)CollaborationismSource codeComputer animation
00:32
Service (economics)FeedbackStructural loadIndependence (probability theory)NP-hardInternet service providerDataflowService (economics)Uniform resource locatorPower (physics)Multiplication signOpen setMappingInheritance (object-oriented programming)Level (video gaming)Mobile appoutputMusical ensembleGoogolBuildingScaling (geometry)MassOrder (biology)SoftwareGoogle Street ViewCASE <Informatik>NavigationBitComputer animation
04:21
SpreadsheetScripting languageRevision controlComputer programmingPoint (geometry)BitField (computer science)Point cloudSpreadsheetScripting languageRow (database)PlanningComputer animation
05:15
Menu (computing)Feasibility studyComputer-generated imageryInflection pointDevice driverComplete metric spaceMaxima and minimaMappingNumberCodeSource codePredictabilityParallel portComplete metric spaceMultilaterationInflection pointMedical imagingComputer programmingField (computer science)Device driverArithmetic meanEndliche ModelltheorieContent (media)Level (video gaming)Limit (category theory)Motion blurTransportation theory (mathematics)Point (geometry)Mobile appView (database)Computer animation
08:14
Computer hardwareDivision (mathematics)Computer animation
08:38
Keilförmige AnordnungPlastikkarteMappingRow (database)Information privacyView (database)Field (computer science)Profil (magazine)Point cloudLevel (video gaming)Goodness of fitScaling (geometry)Duality (mathematics)Medical imagingMusical ensembleUniform resource locatorCoprocessorComputer animation
09:59
Goodness of fitLevel (video gaming)Scaling (geometry)GradientDevice driverInformation privacyInflection point
10:43
Maxima and minimaVideo trackingTask (computing)SoftwareScaling (geometry)State of matterDevice driverInheritance (object-oriented programming)Configuration spaceNeuroinformatikMedical imagingTrailComplete metric spaceSynchronizationIterationPlastikkarteMobile appMachine visionElectronic program guideMotion captureCuboidGame controllerLevel (video gaming)MereologyComputer animation
13:14
Device driverWave packetOperator (mathematics)Endliche ModelltheorieProcess (computing)Computer animationMeeting/Interview
13:53
Computer-generated imageryComputer wormMereologyMappingCollaborationismMeta elementDialectMultiplication signEndliche ModelltheorieBuildingScaling (geometry)Different (Kate Ryan album)Computer animationLecture/Conference
Transcript: English(auto-generated)
00:04
All right. Thank you so much for having us. It's like a huge pleasure to be allowed to present all of the cool work we're doing at Grab. I am head of engineering for the geo department at Grab.
00:23
And yeah, I've been an OSM fan for like 15 plus years. So it's been a long time coming, building maps at a country scale. So for those of you who haven't heard of Grab, we haven't had a chance to see the talk yesterday that we gave. So Grab is a super app in Southeast Asia.
00:42
So we're live in like eight countries covering 650 million people population. So very, very large space, 500 plus cities, more than 25 million users every month. And location is like really key to everything that we're doing. So we offer anything from like transport, food delivery,
01:00
groceries deliveries, but also like financial services and many, many other things. So we are like really like integral to like everything that every time that like some user opens the Grab app, we're powering that with location services. To give you a sense of the scale, like our location team is powering like more than 800 billion API requests a month. So it's like very, very massive scale
01:22
and something that's very critical to what we do. And that's like what the story is today about, right? Like how can we really map and power like all of our eight countries with our own inbuilt solutions, built on house built solutions on top of OSM. So that's really the challenge that we went through over the last few years. And we hope that we can encourage a lot of you
01:42
to do similar things like in your countries and in your companies to achieve that. So, and like to share like open street map at Grab is like been really integral to a lot of things that we've been doing. So right now, 23% of all OSM edits in Southeast Asia are done by a thousand plus people like editing team.
02:03
200,000 plus kilometers of missing roads are mapped. And like every time like user books like a ride or food delivery on Grab, like every time we see that anything is deviating, we're like giving that data back to the OSM community, the road network and improving it. So that was there for a long, long time. But we still like two, three years ago,
02:21
we still used a lot of like external data to power our services. And that was like kind of the journey that we went through. Like, and I mean, I think for all of you who have like tried really like having fully independent maps are like is like doable, but it's extremely hard, right? Like, and there's only like a handful of companies like Apple, Google has like really achieved this at their full scale.
02:40
And they've invested billions of dollars. And while Grab is not like a small company by any means, it's like still a massive challenge. And like, I mean, our CEO and our founders are still like, hey, you don't have like unlimited budgets, you need to do this like really, really cheap. So we're talking, I mean, obviously always, right? Like, I mean, if you're building maps, your CFO is always like in your neck, right? Like you all know that, I guess.
03:01
So when we started building that, I mean, we really started with this vision, how can we be independent in our eight countries? And we started with like OSM at the core. We started like from the foundation that really like map making has drastically changed over the last years. We really built this based on like imagery and AI. And yeah, and we were like now,
03:21
by this time we're like in six out of our eight countries, we were like independent and power maps fully. And then the remaining two will be like very, very shortly. And we're gonna talk to you a little bit through this journey that we took, right? On our case, like what it means to have maps is like two primary things, right? Like the one is like really high quality navigable roads. So it means anything from like building footprints,
03:41
but also like routing, ETAs, all of those use cases. And on the other side, PIs and addresses, right? Like that's the two most important things that we need. And in order to create that for maps, really we figured out like Street View imagery collection is the most critical input data. And then like combining this with AI to extract the stuff, right?
04:00
Like so we've seen that Street View imagery collection is extremely reliable, right? Like you can extract all the data. It's very predictable. You can like see almost like everything on this imagery. And the great thing is also now it's like fairly low budget, right? Like we don't need these massive, like $200,000 mapping bands anymore, but we can do it like with a much, much lower cost scale. And that's what Alex is gonna talk a little about,
04:23
like how our journey for this started and how we arrived where we are today. All right, so the first version of this program was a bit messy if you look at the end-to-end process, but also really effective if we look at the outcomes.
04:41
So how we started is basically go to stores, buy a bunch of GoPro cameras. We hired a few local vendors at that point to run collection. And we managed everything with things like scripts and spreadsheets and a bit of QGIS for planning.
05:00
But basically field collectors would go around, record the imagery, bring the cameras in, copy the data, stitch it on laptops, and then do some minimal QC and then upload everything to the cloud. And this was not like a small pilot that,
05:20
and we changed things immediately, but we actually scaled this across Southeast Asia. We captured most of Southeast Asia with this approach. And yeah, it worked really well in the end for us. Not everything was super smooth, but we did manage to capture a lot of data and we quickly saw how this became
05:42
the number one data source for us, both for roads and especially POIs. And we had the advantage of learning really quickly and providing value right away to Grab and to our map-making efforts. The second thing that we tried in parallel with the GoPro program was just using phones.
06:02
And the idea was instead of buying cameras and working with vendors, can we just incentivize Grab drivers to collect imagery with the CartaView app, which was just recently acquired by Grab at that point. And then we just started that. So basically the drivers, while taking transport or delivery jobs,
06:21
they would mount the phone either on the car windshield, or if they were driving a scooter on the chest mount or on the handlebar, and just run imagery collection and upload everything from the app. So we ran this for a while. We learned a few things. On the bright side, drivers love this program.
06:41
So it was basically like a win-win, right? So they would make some extra money while improving the maps that they were working with every day. And we also proved that doing two-wheel, two-wheel is our code name for scooters, is quite feasible, both the handlebar and the chest mount model. Now on the not-so-bright side,
07:01
first of all, we saw high redundancy, meaning that when drivers are taking delivery or transportation jobs, they often go primarily on the big streets, the main streets of the cities, and you can't have high predictability on the minor roads of each street. And we're not covering that really fast.
07:22
Secondly, quality was sometimes great, sometimes so-so. And also the amount of POIs that we would capture was fairly low. And this is basically caused by the limited field of view that the phones typically have. And yeah, it's more front-facing content than what you see on the sides of the streets.
07:42
And we tried things like lateral kind of mounting. That didn't work because motion blur and rolling shutters and all that fun stuff. So we just basically took a good look at the end-to-end process, and we aim to improve a few things regarding image quality, positioning,
08:02
the amount of POIs, user experience, but also having predictable pace and high completion when you look at the city level coverage. And Filip will walk us through more of the solution for all this. Yeah, so the solution of all of this was that thankfully one day my former boss, our CTO,
08:24
said, hey, Filip, you definitely want to do something better than that. Do you wanna run our hardware division at Grab? I said, sure, hardware sounds like fun. I've never done it, but I'd love to run a hardware division as well. And so what we decided to do is we decided to build our own cameras. Because we've seen that everything that exists
08:42
in the market was either super expensive, we couldn't afford a $100,000, $200,000 mapping run. We couldn't afford a $20,000, $30,000 mapping camera and run them at scale in our emerging markets, which just have not worked for us. So we decided to build our own camera. And that's something that we're really, really proud of. So our team in Shenzhen really built this ground up.
09:03
So we call this Cartacam. You can see them at our booth later if you want. So the cameras are basically purpose-built for map making. So they're connected, so they have a 4G chip in there so they can upload while they're driving. They can take the imagery at home, upload it on a Wi-Fi. It has a high resolution, good field of view. It's built for all day battery life,
09:20
much better than a GoPro. So it has two and a half hours active recording. But we can also trigger it to say, hey, this roads, we have already a lot of images. We don't need to record here. So it's basically can record the whole, operate the whole day and record as needed. It has a two teraops AI processor in there so you can run detection on the edge. Privacy is super important for us to be like blur people's faces on the edge
09:41
that it never reaches our cloud. It has like a dual band GPS chip in there so it's like really high quality location data, not like GoPros that are all jerky, like giving you like really crappy data. And it's like really like cost effective, right? Like it's like the size of a GoPro and just like the cost profile is like fast. It's like very, very similar. So we did that and we deployed that, right?
10:01
Like so you can like see, like we put this like on like driver's helmets. Like our region, as Alex has outlined, is very much like motorcycles, right? Like so we need to have a setup that can go like on bicycles, motorcycles, and narrow roads. So we deployed that and got like really, really good imagery and got like a really good start and we were like, wow, this is like really amazing. Like building like this integrated end-to-end solution
10:21
is really the way to go to like map at scale at a good cost level. You see like some of the AI detections in here, like flake face blurring, license plate blurring. So like this is cool. This allows us to like protect privacy at one side and the other side like collect data. And then we figured out like the one big thing that's like still missing from like really professional grade collection, it's not 360, right?
10:41
Like we wanted to not only map roads, but also POIs. So like in typical grab fashion, our team hacked like a solution. So what better than just put like four of those in a helmet? You need to like go a lot to the gym to be able to carry one of those with your helmet. It's like really heavy. But like in any case, like it demonstrated that it's like a really powerful solution. But obviously we knew that it's not like helmet mountable,
11:00
but it gives you like great data, right? Like so now suddenly we get like super high quality data at like a 20th of the cost of a professional map making camera. So really amazing. And what we are building now is like basically said like, okay, cool, let's like really like build the next iteration of that. So we put them in like an amount and like very ingenious like built for like emerging markets, right? Like you see on this, like there's like this little
11:21
safe box so the driver can like mount the cameras in there, like lock them that nobody steals them. We have like our walkers, walking around with these cameras. So we've really scaled up imagery collection with our like Cardacam 360 light, how we call it and done that. And then the other big part that Alex is gonna, sorry, that Alex is gonna talk about
11:41
is like how we manage the collection. All right, so besides the great images, right? How do we do we collect in older streets in a predictable and high completion fashion? And our answer was basically building great collection software that integrates with the cameras.
12:00
So we call our tool Jarvis. And it does a couple of things for us. So first thing it manages work. So it takes a big city or a county or a state and it breaks it down into like manageable collection tasks. Secondly, we have live tracking of where the cameras are and what they're doing. And we can even remotely configure them, trigger uploading and all that fun stuff.
12:21
It unlocks a lot of possibilities when cameras have sync cards in them. Three, it's sort of like a marketplace for drivers. So drivers see in the app, these available tasks around them. They see upfront pricing for them, how much we pay for them and they just assign tasks to themselves. And then the app guides them where to go
12:41
and what to collect basically with the CartaCam which is also paired to the app. So you don't need to deal with like physical buttons, taking the camera out and doing all that. It's configurable from the app basically. And lastly, it helps us do quality control at scale. So it checks for image quality,
13:00
things like rain or bad lighting with computer vision. And then it checks also for road completion. So we make sure that once the driver says, hey, I'm done with this task, all the roads in scope are captured. So last thing from my side, I just wanna say this is a hyper local operation for us.
13:21
We intentionally invest in upskilling, training some of the local drivers in all the cities in Southeast Asia, instead of hiring vendors and running all that operation. And we actually see that the drivers are doing a better job than the previous vendors
13:41
because the tool helps them to do that and they have good intentions to do that. So it's a highly successful model for us and this is what we're scaling massively in 2022. And Filip will walk us through some of the closing notes. Yeah, so I mean, just to wrap up, we really believe that this model of gig mappers,
14:03
where we can create an income for people creating maps. So I think for a lot of people in our community, it's a very meaningful additional income that they had over COVID in a time that our regions got hit really, really hard. So we really believe in that we can create together with these gig mappers, fantastic maps. And we really believe that it's a great partnership
14:22
with the OSM communities, which we work very tightly with. So we have substantially contributed to make the maps better in Southeast Asia. We've contributed our data and collaborated with lots of other orgs. You have heard our talk yesterday with Meta. So we're really in this where we think it's not like a grab thing, but it's really like a community and a collaboration across much companies.
14:42
So you've seen our OSM contributions. We have Cartier View as a tool available for the community to use. And yeah, I mean, we are really firmly believing in partnerships. We are now working with a few other orgs across different regions to replicate this model. And I think what I wanna leave you with, right, like building maps at like the scale of countries
15:01
and like whole markets to replicate like and replace like external vendors. It's really hard, but it's very, very doable. So we hope that you guys can like convince your like own companies to do more of that and build all of that on top of OSM. Thank you very much.