Follow BigDATAwire:
June 29, 2018

Microsoft Transfers U.S. Map Data to GitHub

Staff report

Microsoft, which reaffirmed its commitment to open source software development with its blockbuster acquisition of the GitHub project collaboration platform earlier this month, is releasing a huge trove of street mapping image to GitHub.

The company announced this week on its Bing search engine blog that it is releasing as “open data” nearly 125 million building footprint images from cities across the U.S. The overhead images of buildings were released to the OpenStreetMap developer community.

Bing researchers gathered the images by combining Microsoft’s deep learning and computer vision tools with an open source deep-learning toolkit called CNTK. The Microsoft development toolkit was then applied to its deep neural network framework along with the ResNet 34, the pre-trained model for PyTorch, to detect building footprints in Bing imagery in a process dubbed “semantic segmentation.”

Source: Microsoft

A second step applied a “polygonization” algorithm used to remove noise and faulty data while detecting building edges to create a map of building footprints.

The researchers extracted 124,885,597 computer-generated building footprints across all 50 states. The open source contribution nearly quadruples the number of building footprint images currently available in the OpenStreetMap repository. Most came from editor contributions or imports from city or county maps.

The U.S. building footprint repository is available on GitHub, as is the CNTK toolkit.

Microsoft (NASDAQ: MSFT), which announced its $7.5 billion acquisition of GitHub on June 4, stressed at the time it would seek to create a balance between maintaining an open development platform where developers could tap into programming languages and tools while “we bring Microsoft’s developer tools and services to new audiences.”

Recent items:

Esri Adds an Unstructured Location Data Capability

Geospatial Data Moves to SAP HANA

Microsoft Acquires AI Specialist Bonsai

BigDATAwire