[OSM-dev] [GSoC] Automatic Street-Sign Detection and Reading - was: [GSoC] A GPX photo stamper.

Stefan de Konink stefan at konink.de
Fri Apr 3 20:24:04 BST 2009

Tijs Zwinkels wrote:
> On Fri, Apr 3, 2009 at 8:51 PM, Stefan de Konink <stefan at konink.de 
> <mailto:stefan at konink.de>> wrote:
>     I would go outside JOSM, and make this a webapp that can be included
>     as a layer in any editor. WMS like.
> I'm not sure whether this is very suitable for a webapp since it might 
> require quite some computation, and will generally operate on databases 
> of pictures that might be a pain to upload. It's certainly possible 
> though. Maybe we should think about a platform which allows it to be 
> used both locally and as a webapp.

sift.openstreetphoto.org already has a bunch of usercollected images, 
all rendered streetsigns from wikipedia. The thing why this should be a 
client side app is the processing might be actually faster (doubtful on 
a mobile though).

>      But please keep in mind that SIFT will *not* work for the typical
>     streetsign because SIFT looks for features, which are typically not
>     present in the 'easy to understand' signs (because they use colours).
> Which.. I of course read minutes after the application deadline 
> expiring. ;) - No sweat, I'll read up on the subject a bit more and keep 
> on adding new ideas to the proposal on the wiki.

You added histograms so I was happy ;) You really want to read 'Indexing 
Via Color Histograms' Swain/Ballard.

> SIFT *might* be able to do some of it's work on features typically 
> present in street-signs. For example: Many street signs have a white bar 
> around them.

And that is exactly what is failing (for me). What actually works is 
recognizing of a bicycle.

> Otherwise we will have to trust on the color-histogram matching, and 
> I'll think about other feature detectors to use as well. (suggestions 
> welcome ofcourse)

As provided ;) See also my previous comment about the sharing visual 
features paper. The sure do magic, especially related to the shrinking 
of the image.


More information about the dev mailing list