Mapillary - Nordugrid

Peter Neubauer

What is Mapillary?

  • A crowd-sourced approach to automatic extraction of high quality data from ground truth

  • Started late 2013

  • 110k contributors

  • 530M+ images

  • Open Data for OpenStreetMap, Disaster recovery, research

  • Opening key components and research

How does Mapillary work?

Capture

  • Phones

  • Action Cams

  • Professional rigs

  • Upload scripts, web uploader, apps

Processing

  • Semantic segmentation via deep learning

  • 3D reconstruction from 2D information

  • Object extraction

Segmented point cloud with objects

Structure from motion

Le Mans - images

Point cloud

Le Mans by paulinus on Sketchfab

Semantic segmentation

Time travel via the point cloud

All-camera support

Deployement at Mapillary

  • Mesos→Marathon→Docker

  • Auto scaling using tiers and spot instances

  • Monitoring using Riemann, Influx, ELK-stack

Integration/System testing at Mapillary

  • Dependency graph via repository structure JSON files

  • Generation of system- and integration test docker compose files

  • Mounting of source resources for easy incremental work

  • Testing framework in external container in Python

  • Analytics and rules enforcing by graph queries

Future directions

  • Targeted capture (capture projects and fleets)

  • Minimal cost routing (Chinese postman problem)

  • on-device processing and pushing computing to the edge