Paige and I have a small project called ImageWiki which lets you
search for images that are similar to other images. We find it
interesting because searching for images is political. "In the future"
our Augmented Reality views will translate images into extended
meta-data and it seems important to have us all own that meta-data
rather than any one company.

The problem is that my implementation is little more than a proof of
concept and so it doesn't articulate the thesis very well. It doesn't
scale, it is slow, it is unstable. The source is on
http://github.com/anselm/imagewiki but ... basically it just isn't
where I want it to be yet... and I am just too busy to bake a solution
from scratch...

Key Indexing for large numbers of dimensions sucks basically. And I'm
looking for advice - actually ideally just something I can slap in and
replace what I have.

The current image recognition implementation is based on SIFT - from
an off the shelf piece of code at the bottom of this page:

 http://en.wikipedia.org/wiki/Scale-invariant_feature_transform

I'm thinking of switching to SURF ( and still need to research how it
stores keys...  ) but I think the keys may be smaller and more
tractable there.

At Where 2.0 I spoke to Blaise from Photosynth and he suggested
talking to two people: David Nister ( http://www.vis.uky.edu/~dnister/
) and Josef Sivec ( http://www.robots.ox.ac.uk/~josef/ ). Other people
to talk to would be appreciated as well.

If anybody has any simple practical advice that is low-energy on my
part ( ideally off the shelf ) I'd like to swap out either our
indexing system for our current SIFT keys or entirely replace the
approach with a different kind of image matching technology. This
would let this project actually work decently and I'd feel a lot
better about talking about it and taking the whole idea further.

Suggestions appreciated,

 - @anselm 415 215 4856 http://twitter.com/anselm

_______________________________________________
Geowanking mailing list
[email protected]
http://geowanking.org/mailman/listinfo/geowanking_geowanking.org

Reply via email to