alt Hacker NewsNo the tech doesn’t work like that AFAIK. The most common use case is exactly localization (think “HD maps” for autonomous cars).
It almost 1-1 data correlation, n-phone Pokémon go scans of a location helping a drone locate itself in the same location in correlation with Maxar’s satellite data.
There maybe some hyper corner case uses. Maybe the billion scans in New York City help them generalize across different phone lenses characteristics, but phone and drone lenses are so different.
Would love to hear some specifics if I’m wrong here.
Replies
> It almost 1-1 data correlation, n-phone Pokémon go scans of a location helping a drone locate itself in the same location in correlation with Maxar’s satellite data.
The headline, which I do understand is in question, talks about training, not using the scans as a database. It is likely that you are right that the scans are not being used to provide localization data, but that is also not what the headline is pointing to.
The headline specifically speaks to using the scans for training. While I do not have any inside baseball, the problem space is often solved using neural nets and other machine learning algorithms. On the surface it seems likely that they would benefit from training data that doesn't necessarily need to be from where the conflict is actually taking place. A base world model, for example, can be developed from data collected anywhere in the world. Its is not an entirely different universe when you step into another country.
But you are suggesting that the algorithms used are entirely classical (i.e. no AI/ML)?
You are creating a 3D model when you scan using Pokémon Go. Difference in lenses doesn't matter, that only matters for the scanning step.
Your hypothesis is correct.
there was a startup that pitched the idea of using Satellite data to do ground based navigation. (https://sturfee.com/vps) they didn't get bought out by either google, niantic or facebook, so it can't of worked that well.
Niantic's stuff is a pre-built map that the client will reference to get a position. Its essentially a massive feature matching exercise. The problem with using airborn photos is that you miss a bunch of features you can't see. (samy thing trying to match ground features from the air.)
THe lens calibration issue isn't actually that much of a problem _for the client_. if you have a rough idea of the lens (exif data really helps there) then you can still get meter accurate (and a few degrees heading) its a bit more of a problem for generating the initial map, but Structure from motion with good motion priors goes a long way to make it less of a problem
Now, Niantic are proposing that you can train a model that can relocalize generally without a detailed map, I think thats a bit far fetch, especially to do at any large scale. (ie bigger than a cubic kilometer)