[pkh80]

Parsing isn't the hard part, its the source data, which if you do the math on what Google has done (drive around the world taking 360 video and LIDAR of streets) is literally billions of dollars worth of work.

TIGER is a pretty bad starting point, geocoding based on block faces is really inaccurate if you want to zoom in to the street level. And its U.S. only.

OSM Nominatim should be a better place to start.

I'd love to see open sourced Street View data collection / processing as part of the OSM project. Then there is a chance to compete with Google.

[ericd]

What you're talking about (massive ground-level driving effort to pinpoint where along streets specific addresses are located) would boost accuracy. Without a Google level address parser, though, you don't get usability for a lot of use cases, which is frankly much more important for a lot of companies. One of the best things about Google's geocoder is that you can throw various location names, as humans type them, and Google will return something, and it's usually the right thing. For many applications, this is the desired behavior, rather than precision.