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

Stefan de Konink stefan at konink.de
Thu Apr 9 13:21:49 BST 2009

On Thu, 9 Apr 2009, Matt Amos wrote:

> On Thu, Apr 9, 2009 at 11:37 AM, Stefan de Konink <stefan at konink.de> wrote:
> > On Thu, 9 Apr 2009, Matt Amos wrote:
> >> On Thu, Apr 9, 2009 at 1:05 AM, Tijs Zwinkels
> >> <openStreetMap at tumblecow.net> wrote:
> >> > I mean, come on, one of these methods must be able to detect street signs.
> >> > :)
> >>
> >> i stumbled across a paper about tracking text in videos[1]. maybe some
> >> of the methods they're using are worth looking at and/or extending to
> >> symbolic signs?
> >
> > Text is again siftable because it is complex. But we are talking with
> > signs (the basic ones) in shapes/colours.
> they are only using SIFT for tracking. initial detection is done by a
> (monochrome) connected components analysis. this might be a useful
> tool for detecting street signs, as they generally come in a small
> number of untextured colours.

Please review the url's i've send before. It can also be exemplar based. I
just heard some other hypeword 'coloursift', but I have to look what the

> afaict, the only non-siftable street sign is the circlular
> white-with-red-border, right?

The one way, preference crossing, will probably also fail. Histograms is
the way to go on those.

> from a practical point of view, what
> %age of road sign instances are that sign? maybe its worth focussing
> on the large majority which are siftable and leaving the "grand
> unified street sign detector" for a follow-up project.

First we need recall, that means a streetsign *detector* then we find out
what it is.


More information about the dev mailing list