Researchers Are Proposing a New Way To Generate Street Addresses by Extracting Roads From Satellite Images (technologyreview.com)
An estimated 4 billion people in the world lack a physical address. Researchers at the MIT Media Lab and Facebook are now proposing a new way to address the unaddressed: with machine learning. From a report: The team first trained a deep-learning algorithm to extract the road pixels from satellite images. Another algorithm connected the pixels together into a road network. The system analyzed the density and shape of the roads to segment the network into different communities, and the densest cluster was labeled as the city center. The regions around the city center were divided into north, south, east, and west quadrants, and streets were numbered and lettered according to their orientation and distance from the center.
When they compared their final results with a random sample of unmapped regions whose streets had been labeled manually, their approach successfully addressed more than 80% of the populated areas, improving coverage compared with Google Maps or OpenStreetMaps. This isn't the only way to automate the creation of addresses. The organization what3words generates a unique three-word combination for every 3-by-3-meter square on a global grid. The scheme has already been adopted in regions of South Africa, Turkey, and Mongolia by national package delivery services, local hospitals, and regional security teams. But Ilke Demir, a researcher at Facebook and one of the creators of the new system, says its main advantage is that it follows existing road topology and helps residents understand how two addresses relate to one another.
When they compared their final results with a random sample of unmapped regions whose streets had been labeled manually, their approach successfully addressed more than 80% of the populated areas, improving coverage compared with Google Maps or OpenStreetMaps. This isn't the only way to automate the creation of addresses. The organization what3words generates a unique three-word combination for every 3-by-3-meter square on a global grid. The scheme has already been adopted in regions of South Africa, Turkey, and Mongolia by national package delivery services, local hospitals, and regional security teams. But Ilke Demir, a researcher at Facebook and one of the creators of the new system, says its main advantage is that it follows existing road topology and helps residents understand how two addresses relate to one another.
Everyone with a fixed "residence" - even if it's not on any road - has an ICBM address, even if they don't know it.
They also have other "your location is your address" addresses.
Knowledge is how to play a game, intelligence is how to win, wisdom is knowing what game to play.
Consider what a generative adversarial network (GAN) can do to such an algorithm. Perhaps all an attacker needs to do is paint some rooftops in a manner that creates an off-by-one error on counting streets and then they can misdirect (or even intercept) somebody's mail.
Read this article on Camouflaged Graffiti on Road Signs Can Fool Machine Learning Models (from 1y ago) for some examples of how simple stickers or graffiti can fool these systems. It also hit Slashdot, though all I can find is an article on 'Psychadelic' Stickers That Confuse AI Image Recognition.
Use my userscript to add story images to Slashdot. There's no going back.
I like Plus Codes. They are well thought out and open and free. I'd like to see more adoption of it.
In a lot of places in the world, there is no existing method for having addresses. Many people aren't interested. For example: Costa Rica Street Signs and Addresses. “ We don’t need no stinking street sign!”
Personally, I'd think it might be just as well to incorporate high resolution Lat-Lon coordinates into a QR splotch and let local software on a smartphone find the route. Addresses with words are designed for people to learn them, but people don't learn them. We use address books.
In most places, you can't use satellite images to trace roads because of this minor obstacle known as TREES! Yeah, I've looked at Google Maps in satellite mode.. a lot.
I've abandoned my search for truth; now I'm just looking for some useful delusions.
Because most people understand their location relative to a road, trail or intersection. The problem isn't to assign an identifier to each location in a town. It's to get people to remember and use something that they are already familiar with.
Have gnu, will travel.
https://plus.codes/ is the URL.
They work in Google Maps, BTW.
There are rural places all over the world that need a good addressing system that works regardless of internet connections.
They are open. Anyone can use the algorithm. That's why W3W fails with their proprietary crap. If w3w opened it up, it is a great tool.
Plus.codes let you determine the resolution - need a few acres or a tiny food stall - you decide.
They don't use look-alike letters/numbers.
I lived at 66XQXRXV+XC for a few months, then moved to 762Q858M+9G. The first official address was "500m south and 500m east of the local church". The 2nd official address was "500m north of Johnny's Pizzeria" .. but Johnny's had burned down 5 yrs before I moved there.
In Nepal, I lived at 7MV7P9MG+9P. To get snailmail, packages are mailed to the local Ma-Pa convenience store a few blocks away. They want a cell phone number on the outside so the manager can call it to get the package picked up.
There are lots of places in the world that don't have addresses - Nepal is one. Costa Rica outside San Jose is another.
Tokyo addresses are numbered in the order they were built, so #1 could be 3 miles for #2 and #67 could be next to #2.
South Korea is using 2 addressing methods now - the old one and the newer one which is more like the USA. It will take another 30 years for the older addresses to stop being used - need people to die off. Plus Korea has included delivery in many of their stores for 50+ yrs.
PlusCodes work in OSM tools as well, just not all of them, which is too bad.
Why not just use GPS coords? Because there are 3 different systems (dec, hms, h.dec) and people don't memorize 14 numbers very well.
There are a few other methods. None are as clear, simple, concise, as plus.codes.
The implementation code was in github last time I looked.
There's probably arguments that you don't need addresses at all, but I think the arguments for non-street addresses are nerd masturbation. What they totally fail to grasp is that in most cases a local reference is sufficient, for an analogy with date and time "Tuesday at two" is usually meaningful in context, you don't have to go ISO 8601 and say 2018-12-04T14:00:00 because it's obvious you mean this year and the coming Tuesday and two in the afternoon not two in the middle of the night and on the strike of the hour. Unless it's not, in which case you append as necessary. Same thing with global coordinates, you don't need a globally unique address to deliver me a pizza. I don't want an address where a lisp or typo sends me halfway across the world. I don't want to constantly repeat "planet Earth, Solar system, Milky Way galaxy". The "old way" was way, way more fault-tolerant .
Live today, because you never know what tomorrow brings