The MapAI competition proposed two different tasks to segment buildings, where the first task only utilized aerial images, while the second depended on laser data (LiDAR) with or without aerial images.
The participants had to create a program - a machine learning algorithm - where artificial intelligence created map data or building data from an aerial image and a laser data set, possibly in combination with an aerial photograph. A released training and validation set with data from Denmark was used in the competition.
The winner's podium
- Team Founder: Eivind Hovdegård Furdal and Lars Martin Hodne. Total score: 76.35%
- HVL-ML: Sathiesh Kaliyugarasan. Total score: 75.85%
- UIAI: Lukasz Filip Mrozik, Andreas Høiberg Eike and Pedro Alves. Total score: 74.37%
The winning team of the competition won 1,200 euros, second place 500 euros, and third place 300 euros.
The results show that it is possible to automate processes to create precise map and building data using machine learning.
- The competition took Eivind Hovdegård Furdal and me into quite new terrain, and we learned a lot along the way, says Lars Martin. They have shared the knowledge internally in Norconsult Information System, where they are both employed.
The Norwegian Mapping Authority organized the MapAI competition in collaboration with NORA (Norwegian Artificial Intelligence Research Consortium), AI:Hub, CAIR (Centre for Artificial Intelligence Research at the University of Agder), Norkart and the Board for Data Supply and Infrastructure in Denmark.
Great value for doctoral study
- These are very good results! The competition had difficult conditions as the training data is based on true orthophotos, and the test data is based on "ordinary" orthophotos, says Sander Jyhne, PhD student at the Norwegian Mapping Authority.
Jyhne will use the competition results in her doctoral thesis.
Building segmentation competition
The competition started with 36 teams. In the end, eleven teams delivered code for the tasks, of which nine teams also wrote an article about the method and results. The submitted models were tested against an unpublished data set to see which one interprets the data best.
The competition was divided into tasks:
- Segmentation of buildings only when using an aerial view.
- Segmentation of buildings using Lidar (laser data) with or without aerial photography.
To win, the participants had to complete both tasks and submit a two-page article on method and result. A simple code had been set up for data retrieval, training and evaluation, so the bar was set relatively low for interested teams who wanted to participate.
Extra curious?
At the Nordic Machine Intelligence Journal you can read in more detail about the competition, find the formula for obtaining scores and other characteristics about the data set and the tasks.
Article by: The Norwegian Mapping Authority (Kartverket).
The article was first published by the Norwegian Mapping Authority in Norwegian. To access the original article, click here. NORA gained permission to use and translate the article.
This articles was originally published on the 30th of January, 2023.