Catt (with Cope)

Talking about going places on Flickr. One specific problem. Want to be able to say where photos are taken. We have this “where on earth” database that has lots of places in it. Really good because people may search for places by different names even if the named areas overlap a lot. It’s not so good for reverse geocoding as you might end up saying the wrong thing. This is really hard. Reverse geocoding gets a lot fewer results from a google search..

At ETech in 2007, photos taken would be said they were at San Diego County Jail.

If people are going to take photos and go to the trouble to geotag them, we should be able to describe the places better.

Reverse Geocoding

“Nearest linear object”

We tried to reduce the type of places to a base set of commonalities. First assume at street level, then go to neighborhood, then locality, county, region, country. About a year and a half into this we’ve decided to go with street, locality and airports, “metro”, …

FireEagle uses something very similar. FE have adjusted their model for privacy regions (if you’re in a tiny town, you’re obvious).

Dopplr has a different model. They go by the distance between San Francisco and San Jose. This is all they need.

Geonames, great, gets it wrong sometimes too.

“What’s going on?”

“Imperfectly transmitted code”

People have different ideas of what things mean.

We work with bounding boxes.

I geotagged a photo in Millerton state park, it said it was in “Inverness” which is on the other side of the water.

Showing a slide with bounding boxes. They intersect. Do lots of iterations to filter them out. We should end up seeing that milerton is local and inverness less so. We adjust the measurements on other parameters, but this one was just wrong. We geocoded Petaluma and fixed it.

We take 78 steps to go through and figure out where this could be.

We have a responsibility to do useful things with our users data. So we’re asking for help. In a few weeks time we’re going to ask “Is this right? If not let us know.”

If people keep telling us that things are not where we thought they were, we can take that data back in and start fixing things. This will hopefully roll back into the system and give us more precise data. Like a beach gives us an idea where the coastlines are.

We’ve spent huge amount of effort trying to get this right.

Going Places on Flickr

Technorati tags: flickr, geocoding, reverse-geocoding, where, where2.0, where2008