Comment by peterburkimsher | Hacker News Reader

Comment by peterburkimsher | Hacker News Reader

peterburkimsher

tl;dr - Google Maps has buildings and Areas of Interest (streets with more businesses) based on satellite & street view data.

"Google is creating data out of data ... It makes you wonder how long back Google was planning all of this—and what it’s planning next"

My guess is that they weren't planning it. Somebody decided to use their 20% time to learn TensorFlow and process the Street View imagery for fun, and that side project got promoted into Maps.

When I downloaded 10 years of church notice sheets, I was trying to get the lyrics of the songs in Chinese. I later realised that I could also find the most common song names, and then focus my Chinese-learning on those songs. Big data is all about gathering more data than you know what to do with, and figuring it all out as you go.

This was definitely planned, only perhaps not in this kind of detail. The Geo organization's motto paralleled the company's (to organize the world's information, etc. etc.), only at a geospatial level.

The data you see in Google Maps came and probably still comes from a pipeline of pipelines that rebuild the whole planet on a regular schedule. Since we're talking about all of Earth, random stuff here and there could cause hiccups and delays. Information comes from all sorts of sources, including custom HW (streetview camera, etc.), flying their own planes, owning a satellite company for a few years, and so on.

In one shape or another, it all predates Tensorflow by years.

I reckon Google's 3D building designer, that allowed people to plot buildings and place them on the map was used to train an AI to create buildings from features ... maybe. The page says satellite data, matched with street level, but I often see planes flying over my UK city and imagine they could do radar or similar tracking to match with visual data.

I do not get what makes this post so popular. Google Maps is often better and uses automatically extracted features from photogrammetry. Eh. 350 points at most.

Reply on news.ycombinator.com