Hacker News Reader

Comment by maxglute | Hacker News Reader

I plugged in satelite view of pacific islands and it's giving random guesses. I used some of the source images listed on the bottom, one's with properly labelled island name, and it's still giving random guesses.

Random streetview shots seems to guess countries correctly so far, or at least region with similar construction. But that still means being (confidently) off by 1000s of km.

View on HN

Speaking of Street View, the person who stalked a Japanese idol through her eye reflections [0] also used Street View.

[0] https://www.nbcnews.com/news/world/stalking-suspect-allegedl...

Satellite view? This is about street-level photos… I don't think you would mix these two things in the same model, it'd just dilute the capabilities on both.

It identified aerial shots of hawaiian islands, prince edward island, cuba and most of the commonly "known" islands with distinct shapes fine. At least much more consistently than random street level views I've thrown at it, where description seems sensible, in as much as a good geoguesser can descripe their rationales.

I don't question that some models can also handle satellite views, just that it makes little sense to "pollute" a model intended for geolocating photos with input data from a different category. Especially since satellite views are just vaguely similar enough to photos that I'd assume they can cause quite a bit of unnecessary ambiguity and model complexity to handle it.

> I don't think you would mix these two things in the same model, it'd just dilute the capabilities on both.

I'm not sure. Isn't the consensus that neural networks get better, if you make them bigger and give them harder problems?

> if you make them bigger and give them harder problems?

I mean… maybe… but that also makes them more expensive to run…

Reply on news.ycombinator.com