4 ms·

Seems like an AI could perform this sort of analysis to geolocate images with far fewer distinct features and no reliance on any prior knowledge, since it would

Seems like an AI could perform this sort of analysis to geolocate images with far fewer distinct features and no reliance on any prior knowledge, since it would have near omniscient knowledge of all satellite imagery. For example, this seemingly unidentifiable shot of Gwern’s window view [0] contains a nondescript tree next to a body of water. Nigh impossible for a human without any prior knowledge of Gwern’s general location to identify, but potentially trivial for an AI, since it could automatically scour all satellite imagery for trees and shrubbery of similar shape next to a body of water and a house with a three-sash bay window, a task utterly infeasible to do manually. [0] https://www.gwern.net/Links#hardware https://www.gwern.net/Links#hardware

vanattab 4y ago
I don't work in AI but would this really be as easy as you say. Plants might not seem to change much in the time you are observing them but they can change significantly throughout the season and are constantly "reshaped" by blowing winds and the like.
garbagetime 4y ago
This seems like a possibility. Train an AI to identify and measure features in satellite imagery (water bodies, trees, lawns, buildings, etc.) Allow a user to search for areas of size x m^2, that have water bodies of at least y m^2, forested area of at least z m^2, etc.
aaron695 4y ago
pvaldes 4y ago
Humans are better picking tiny details. My first candidate would be a young oak, maybe an apple. Possible poplar at right. Computers keep generating a cat gravity field around, it seems. Who is Gwern? (or if you prefer: Why are we peeping at their nice workplace?)
ZeroGravitas 4y ago
The Bellingcat book goes through several of these kind of investigations. They also discuss the potential for AI to help. One thing they mention is getting approx time from shadows when you know the location. That photo has a blurry date/time in it. Which lets you use shadows to find approx latitude I think.

vanattab 4y ago

I don't work in AI but would this really be as easy as you say. Plants might not seem to change much in the time you are observing them but they can change significantly throughout the season and are constantly "reshaped" by blowing winds and the like.

garbagetime 4y ago

This seems like a possibility. Train an AI to identify and measure features in satellite imagery (water bodies, trees, lawns, buildings, etc.) Allow a user to search for areas of size x m^2, that have water bodies of at least y m^2, forested area of at least z m^2, etc.

aaron695 4y ago

pvaldes 4y ago

Humans are better picking tiny details. My first candidate would be a young oak, maybe an apple. Possible poplar at right. Computers keep generating a cat gravity field around, it seems. Who is Gwern? (or if you prefer: Why are we peeping at their nice workplace?)

ZeroGravitas 4y ago

The Bellingcat book goes through several of these kind of investigations. They also discuss the potential for AI to help. One thing they mention is getting approx time from shadows when you know the location. That photo has a blurry date/time in it. Which lets you use shadows to find approx latitude I think.