Hacker News Reader

Comment by mooneater | Hacker News Reader

Sounds like simple crowd sourcing. But please note what they are really doing.

They are providing an ideal training set for training machine learning algorithms to automatically extract 3D models from aerial photos.

Yes, so they can replace the crowd with bots.

Edit: spelling.

View on HN

It's true, but since everybody has access to the test images (so they can label them), and the models (through google earth), it seems like anybody can train whatever algorithm they want on the same set.

I think it's forbidden. Check the terms of use for Google Earth imagery and models.

Depending on how much us humans contribute, there might not be much left for the bots.

The problem that they want to solve is: "How do we create a three-dimensional model of every built structure on Earth? How do we make sure it's accurate, that it stays current and that it's useful to everyone who might want to use it?"

I think "stays current" is an important part of the question. Over a period of many years, buildings in an area are taken down and new ones are built.

Maybe Google will have an enthusiastic mass of people contributing to this project, and end up with good models of most of the buildings out there right now... but in ten years, are people still going to be doing this for them? I think mooneater's suggestion makes sense if you look at growing this into something big and robust in the long term.

Yes, keeping the model up to date is definitely an important issue. Don't forget that this also involves capturing new aerial imagery periodically.

Another approach that may have more potential than automatically extracting 3D structure from aerial imagery is using aerial depth maps from something like LIDAR. There's a research group at USC doing this, and you can see a video describing the process here:

http://graphics.usc.edu/~qianyizh/miscs/cvpr09_streaming_ful...

And here's another video from USC that just shows textured 3D models a few cities, captured using LIDAR and aerial imagery:

http://graphics.usc.edu/~luwang/videos/cvpr.avi

Yeah. This is a bit ridiculous, conning people into working for them for free at this project by calling it a game. What is up next - the "improve my search algorithm" game?

I wouldn't be surprised if they already have something in the pipe, seeing as how they already have this:

http://images.google.com/imagelabeler/

They also have been using 1-800-GOOG-411 to train their voice recognition engines (like Voice Search in Android)

I think it's win-win. Google gets data and somebody somewhere is happy to be doing something more productive than FarmVille or some other wacky Facebook game people like you and I avoid like the plauge. Hell, I'd pit building makers against each other by having leaderboards of number of buildings, average building rating on a 5-star scale, etc.

Outsourcing research isn't that crazy: http://www.netflixprize.com/

>What is up next - the "improve my search algorithm" game?

http://www.netflixprize.com/

It's a bit unfair to call it a con. You can choose to try it or not. Likewise you can enjoy doing it or not.

It actually is pretty fascinating. I sat down to test it and ended up modeling four buildings in my downtown area. ;)

It works quite well and it is actually rather nice to see the building go up. It will be even nicer if I get to see them included in Google Earth. We'll see.....

At any rate what I am saying is that I don't mind working for them for free.

Why not, worked for the image service .. I played a few times it was kinda fun in a waste-a-few-minutes way. Pretty good vocab tester, think of different ways to describe a scene, think of ways that your "opponent" is likely to use, think of the most complex way that is still likely enough (more points).

It works both ways... why do you think Google offers free 411 service?

(hint: voicemail transcription)

Reply on news.ycombinator.com