KartAi – An open living lab for Ai in Norway
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 | 351 | |
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 | 10.5446/69211 (DOI) | |
Publisher | ||
Release Date | ||
Language | ||
Production Year | 2022 |
Content Metadata
Subject Area | ||
Genre | ||
Abstract |
| |
Keywords |
FOSS4G Firenze 202234 / 351
1
7
13
22
25
31
33
36
39
41
43
44
46
52
53
55
58
59
60
76
80
93
98
104
108
127
128
133
135
141
142
143
150
151
168
173
176
178
190
196
200
201
202
204
211
219
225
226
236
242
251
258
263
270
284
285
292
00:00
James Waddell Alexander IIDigital signalSlide ruleDecision tree learningLevel (video gaming)Projective planeMeeting/InterviewComputer animation
00:20
James Waddell Alexander IIDigital signalOpen sourceLevel (video gaming)Decision tree learningData managementProjective planeSystem callComputer animation
00:42
QuicksortIntegrated development environmentComputer animation
01:11
RobotCategory of beingWebsiteData storage deviceArchitectureDatabaseArtificial neural networkSource codeObject (grammar)Endliche ModelltheorieEstimationBuildingDigital signalOperations researchTexture mappingProcess (computing)CollaborationismSoftwareProjective planeQuicksortBuildingMereologySpacetimeDigitizingLoop (music)Wave packetElectronic program guideDifferent (Kate Ryan album)Machine visionProcess (computing)Virtual machineCategory of beingCASE <Informatik>Data storage deviceExtension (kinesiology)Multiplication signLevel (video gaming)Square numberDisk read-and-write headCartesian coordinate systemResultantEndliche ModelltheorieUniverse (mathematics)ChatterbotRoboticsLangevin-GleichungInstance (computer science)Musical ensembleData typeMathematical optimizationMetropolitan area networkOpen setComputer fileSimilarity (geometry)Intrusion detection systemMedical imagingMetreDiallyl disulfideCase moddingComputer animation
07:59
Fundamental theorem of algebraUniform resource locatorBuildingExtension (kinesiology)Virtual machineEndliche ModelltheorieSheaf (mathematics)BenchmarkLevel (video gaming)Wave packetDifferent (Kate Ryan album)4 (number)Type theoryCase moddingComputer fontMusical ensembleImage resolutionSoftware testingResultantSoftware developerMedical imagingMereologyMetropolitan area networkTime seriesData storage deviceSet (mathematics)Electric generatorVirtual machineCASE <Informatik>Open setInstance (computer science)BenchmarkAreaDeterminantEndliche ModelltheorieBuildingProcess (computing)Core dumpExtension (kinesiology)Object (grammar)Flow separationStandard deviationDemosceneSoftwareQuicksortComputer architectureBitSpacetimeProjective planePerfect groupPoint (geometry)Task (computing)WordMassComputer animation
14:48
Virtual machineBenchmarkEndliche ModelltheorieBuildingSheaf (mathematics)SoftwareObject (grammar)AreaComputer-generated imagerySystem identificationDatabasePixelInclusion mapPredictionMaß <Mathematik>SoftwareResultantProcess (computing)Goodness of fitSoftware testingDatabasePopulation densityMusical ensembleWave packetBeat (acoustics)Doubling the cubeRevision controlGrass (card game)Level (video gaming)Execution unitCase moddingSoftware frameworkSolid geometryFlow separationCompilation albumSoftware developer1 (number)BuildingProjective planeDifferent (Kate Ryan album)Boundary value problemAreaDiallyl disulfideRegular graphDifferential (mechanical device)Data modelEmailQuicksortCollaborationismVisualization (computer graphics)Universe (mathematics)Endliche ModelltheorieMereologyObject (grammar)Task (computing)Similarity (geometry)Set (mathematics)BenchmarkComputer animation
21:37
BuildingMaß <Mathematik>CodeCollaborationismJames Waddell Alexander IIMedical imagingResultantProjective planeEndliche ModelltheorieReal numberBitMusical ensembleLevel (video gaming)CollaborationismComputer animationMeeting/Interview
Transcript: English(auto-generated)
00:00
Thank you very much. I'm going to talk about, yeah, there's the slide. Talk about Cart AI, which is the Norwegian for Map AI. So it's a Norwegian R&D project.
00:22
I work at Norcart, which is a private IT company working a lot with the GIS, using a lot of open source GIS tooling as well. And I'm the project manager for Cart AI, which I'm going to talk, call Map AI, I think, for now.
00:44
We were inspired by sort of our motivation that we believe that citizen involvement creates better data and better data democracy. And we work a lot with cities and the municipalities in Norway, which are focused around, obviously,
01:00
creating a good environment to be a citizen. And data and data democracy is really important for that. And the whole idea for this R&D project stemmed out of, if we could imagine having a chatbot like this, a robot,
01:23
so to say, that could proactively engage with the citizen, with the property owner, and ask, this could be a chatbot, could be a text message, could be whatever. But you can imagine, be proactively contacted and engaged in the conversation, where
01:41
Carly, the citizen robot, could ask, have you built a garage on your property in 2016 or whenever? And you could answer with text or with a button or something, and you could say that, yeah, that's correct, we actually built a garage
02:01
below 50 square meters, and that's important in Norway because it's allowed in Norway to build, without a permit, you can build below 50 square meters, but you have to report it in, which many of us don't do. So it's a lot of data gaps for the municipalities.
02:22
But further on, the robot could then just do all the casework and update the cadastre, update the map, update the case files, and everything would be okay. And when I present this for the municipality workers,
02:43
for the caseworkers that work with this day-to-day, they're shaking their heads and it's not able to do this. But we think that it could be possible, but it opens up a lot of problems, and that's what the idea of the project is,
03:00
is to address some of these issues. We're not going to solve all of them, and we're not going to make this robot, but it sort of guides us in the project. And I'm going to highlight some of this for this crowd, which is really technical. And that is, one of them are,
03:21
how do we reuse training data in AI, and how do we store the data, what is the storage architecture? When we collect different data types, so that would be aerial images, it could be map data, cadastre data, and also training, optimized training data,
03:42
how to store it and use it on a national scale, which we aren't able to do today. And how can we find building, detect and segment building footprints, and even potentially 3D buildings from this data? So we're using aerial images,
04:02
but we're also working towards using oblique imagery. And how can we, the non-technical part of this is, how can we really create a proactive and fully digital citizen dialogue? And that means putting the citizen
04:20
in the machine learning or AI loop, so to say. And we also have some regulatory issues, which I'm not going to go into today. So that sort of sparked the idea of the R&D project, which is a collaboration between the city of Kristensen, where I live, in the southern of Norway,
04:43
and the Norwegian Mapping Authority, and the NORCART, and the university in Agder. And it's ongoing. We have been working with it a couple of years, but we're going to work further on with it as well.
05:01
And the whole intention of it, and we also received a lot of goodwill and praise around that, is that the intention of the project is that we're open, we're working openly with it, even though we have different partners that are fully commercial and make proprietary software and so on.
05:20
But the whole idea of the project is that we're going to work open, everyone is invited to collaborate and join, but the results need also to be open. So that means that the AI models, the training data, all of the good things that we develop in this project are going to be released openly,
05:41
which is not that common in Norway, for all IT projects, I can say. So our motivation in the project is concentrating around the building application permit process, which is a huge manual process today for all of the municipalities.
06:01
We have around 100,000 building application permits per year in Norway, so you can imagine it's a lot of work and if the data, the ground data, the disaster data, the map data is out of date, for instance, missing a building, missing an extension to the building
06:21
or missing a garage, then the time it takes to do the case work really, really increases. Yeah, you can imagine it costs more, it becomes inefficient. And we also sort of have a stretch goal here, more vision or idea that we really hope
06:43
that the open process and the way we work with this project and the way we share data, the way we open up the AI data and models across the private public sector is going to sort of spark the innovation and inspire more projects and more private public
07:07
collaboration like this. And in Norway, we received a lot of attention around this project because we're working a bit different than the rest, some similar ideas of an R&D project.
07:23
And I think, personally, I think that is because we're connecting the AI, the GIS, the cadastre, which is their own knowledge or their own silos, and so to say, connecting that with citizen dialogue
07:43
and putting a human in the loop kind of thinking. And also that we're working across private public sector and working in the open space. And this sort of illustrates the idea of the project,
08:00
which are sort of said in words. What is the storage architecture combined with the machine learning, the AI models? How can those type of things collaborate? Are there any standards on national and national level that we could use or could develop? And then put it further into the case work,
08:24
into the work process of the municipality, which is not that common today. Today, it's more the software space one develops on the one hand, and then the user, then the user, the caseworker is using it,
08:43
not being part of the development. And we think that maybe it's a bit braggy this, but we believe that we're really, really lucky in Norway that we have sort of a tradition working together to form the geospatial data, data sets in Norway,
09:05
especially around aerial imagery and the map data, which is collected on a national scale at around approximately one year intervals.
09:22
And with really, really high accuracy. So we have the whole of Norway in around 10 centimeter resolution, which are open to sort of R&D projects like this and open for sharing across at least the public sector. So that means that the one small municipality
09:43
or small city could use the whole data set. And we have time series for this for several, several years back, even for 1920s, I think not the whole of Norway, but at least part of it.
10:00
So that leans very well towards doing more stuff like this. And the recent years, say five years or so, we have also been able to generate high accurate 3D models of buildings and of the terrain from map data generated from the aerial imagery.
10:23
And as we saw, many things can be done with point clouds. And I think it's finished now, but just recently, we have LIDAR scanned the whole of Norway. So we have that as well to be put into this data store.
10:45
And we work a lot with GIS nerds. I'm a GIS nerd myself. And when I see this type of illustrations like this, I see a building map, building outline map and an aerial imagery or rectified orthophoto.
11:04
But when we talk with AI developers and bring core AI developers into the project, they don't see the building. They don't know what the building map is. They don't know what the aerial image is. Even they really, really have no idea of photogrammetry
11:22
and the stuff like this. But what they see, it is training data for AI. And that means that we have a massive amount of training data, which we can train the AI models on. And these are some of the really, really first results
11:43
we got for several, several years ago. We're just sort of prototyping, is this possible to use the map data combined with aerial imagery and use it as training data to segment or detect different objects.
12:00
So what we're aiming for here is to generate an AI model or several AI models which could detect one of these scenarios and hopefully segment it as well. So either there's an existing building, there's an extension to the building,
12:20
there's a new building, or it's demolished. These are the four scenarios that we're looking for. So it seems pretty simple, simple task at hand, but believe me, it's not that easy for an AI. It's not that easy for the GIS person in the municipality to detect the differences as well.
12:45
And we have all this accurate map data and have a lot of high resolution aerial imagery. But what we found pretty early in the project is that there are some minor, minor flaws in the map data.
13:01
It's skewed or it's not aligned accurately with the rectified orthophoto, for instance. So what we did was to select different geographical areas, which you can see, yeah, not that good, but you can see it highlighted in the image here.
13:22
And we manually adjusted and made more or less perfect training data in these areas, which are high for Norwegian. These are highly dense areas. And that we collected, we gathered that data,
13:42
we gathered some more data that we have made open for everyone to use. And we also gathered the training data, which are optimized for AI modeling and AI training. And we also did a lot of benchmark data,
14:02
which I will come back to, and a lot of testing with different AI models. All of this are collected and gathered and are going to be made available for others to use in their case. And we're working now, actually, to extend the geographic footprint to release more and more of this data
14:25
for open access. And this, this autumn, we're going to be hosting a competition, an international competition, which is going to be published. This is a real, real draft version, so don't take notes on this.
14:42
But we're going to do the competition, it's going to be released open, and in collaboration with AI researchers at the university. And that's open for not only researchers,
15:01
but for you guys and all others, where we are going to also open up the data sets and data models. So it could be really good if you're interested. I urge you, either come see me or reach me at my email
15:22
or other places where you can find me. I really hope that a lot from the GIS community can collaborate on this and join the competition. Because what we're seeing is that the IT, the software, AI developers are really in need
15:40
of the geospatial conference. So to move on to more of the results that we are seeing, we started off the project using off-the-shelf software. So we used TerraSolid and ArcGIS Pro because they both have built-in features
16:05
for detecting and also segmenting buildings. And it was useful to sort of see what's the benchmark of more or less de facto standard software,
16:22
which is currently a lot of cities, the bigger cities, the agencies are using it today. So the task at hand was more to see, could we beat it? Could we develop an AI model or several AI models
16:42
that could beat the results of ArcGIS and TerraSolid? And we developed our own training model. It's a version of U-Net, which we adapted and adjusted and also made some adjustments to it.
17:03
We call it Ki-Net, which is the name of it. And we trained it on a lot of areas where we have a high density of buildings, or regular buildings in the region sense.
17:21
And that made us, we got a fair amount of results using it. We actually beat ArcGIS as well with higher IOU, Intersection of Union. We also used another score or another metric which is called Boundary Intersection of Union,
17:40
which is what we heard, but it's only the boundary of the object that we're interested in. Because finding, which we're going to see, detecting the building is not that problematic,
18:00
but segmenting it and getting the edge sharpness is really where the hard part is. And that made us test a lot more. And we tested on more classic frameworks and we also tested some really newer ones.
18:23
And we made some tests using the results, I think it was on Bait model, where we double trained. So we made first, we tried to segment from one training
18:40
and we used the results to train another network, which didn't work yet. But we were hoping for the ideas that we need to get the edges even more sharp than we are today. So this is some of the results,
19:02
not in the tabular form, but more visual. What we saw, first of all, is that we used the orthophoto and we used the national map database. And we saw that using the map database directly gave us fair results.
19:23
But using the map database and the adjusted map database, which we call it, the high accurate training data, which is not millions of features, but it's some thousand.
19:42
That gave us significantly better results. So that's a learning experience from this. So we think that it's good value to provide this high accurate data.
20:01
And these are some of the results. Today we are able to get even more accurate, but it illustrates the point of it. And we're able to, let's see here, we can, yep, you can see here, we have a lot of grass,
20:22
grass on the roof in Norway, which is really, really difficult for an AI model to differentiate between what is grass on the ground and what is grass on a roof. And we also have some minor flaws as well.
20:41
But in general, we get really good results. And then we added LiDAR data and height data from LiDAR. And in the training process, obviously we saw a huge difference. So that made the model really, really much stronger,
21:03
the differentiating between patios and grass, grass roofs, and similar. And then we extended our training data set. We sort of forced the training where there were not that many buildings.
21:22
So we retrained and retrained. And we have some difficulties. The hardest one is to differentiate between multiple cadastric units within the same building. So is it one building or several units? Which is sort of a hard problem. These are some images of our results,
21:43
which is good enough, we think, or it could be really much better. But it's pretty good, I would say. And now we're actually deploying some of these models in the real world, helping the caseworker or the GIS person in the city
22:03
to be a bit more efficient. So yeah, this year we're going to deploy some of these models in the real world scenario, which we think is really, really cool to get some value out of the R&D project as well.
22:21
And we're going to continue throughout this year and hopefully get more funding for the next years further on. So I would really urge you to just reach out and contact me if you would like to join the competition, collaborate on the models that we're doing.
22:43
We're really eager to try it on other international level, not only in Norway, to see if it can work. So that's all. Thank you.