Hacker News new | past | comments | ask | show | jobs | submit login

Training data is an issue. Microsoft released some datasets of buildings in North America (125 million footprints). Later, they decided to use the same technology to generate building footprints in Uganda and Tanzania:

https://blogs.bing.com/maps/2019-09/microsoft-releases-18M-b...

They needed different training data to cope with the different landscapes.

Facebook has faced similar challenges trying to use ML to generate road centerlines for OpenStreetMap.






Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: