Mapillary: Computer Vision Crowdsourcing with Peter Neubauer


Manage episode 218626626 series 1437556
By Discovered by Player FM and our community — copyright is owned by the publisher, not Player FM, and audio streamed directly from their servers.

Mapillary is a platform for gathering photos taken by smartphones and using that data to build a 3D model of the world. Mapillary’s model of the world includes labeled objects such as traffic signs, trees, humans, and buildings. This 3D model can be explored much like you can explore Google Street view.

The data set that underlies Mapillary is crowdsourced from volunteer users who are taking pictures from different vantage points. These smartphone photos are uploaded to Mapillary, queued, and processed to constantly update and refine the Mapillary model.

Mapillary processes high volumes of photos from around the world. The images in these photos need to be correctly fit into Mapillary’s model of the world like a puzzle piece sliding into place. The image needs to be segmented into the different entities within, and those entities need to be put through object recognition algorithm. When two pictures have a conflict, that conflict needs to be resolved.

Mapillary is full of interesting engineering problems. The high volume of images and the level of processing has created the need for a unique sequence of indexing, queueing, and distributed processing using Apache Storm. In addition to processing all of this data and building a 3-D model, Mapillary serves an API for querying geolocations about traffic signs, road conditions, and bus stops.

Peter Neubauer is the co-founder of Mapillary, and is also a co-founder of Neo Technology, the company behind Neo4j. Peter is a world-class engineer and he joins the show to give a detailed overview of the technology behind Mapillary, from ingressing the photos to running data engineering jobs to serving the API.

Show Notes

The post Mapillary: Computer Vision Crowdsourcing with Peter Neubauer appeared first on Software Engineering Daily.

135 episodes available. A new episode about every 7 days averaging 58 mins duration .