What is it?
Figga is an image search engine. Instead of entering search words
or keywords you are able to create a "keypicture" or image by
drawing and/or uploading a images of your own. It than takes this search
image and returns a set of visually similar images. It is a bit like Google's
image search but with the fundamental difference that it relies solely
on matching the visual features of images.
The images it returns don't make sense?
Sometimes the images it returns are obviously similar and sometimes
they seem random, mistaken or "false positives". When the software
is searching for matches it is entirely uninfluenced by the meaningful
content of the pictures it is comparing, nor by any kind of object recognition,
file names or textual annotations. It is purely looking for similarities
in their visual features - composition, light and dark, shape and pattern.
If you look carefully you can often see visual elements that match your
keypicture, but in ways that are quite unexpected. As you get used to
understanding this "algorithmic" matching process you will find
that the software is teaching you to "see" images in the same
that it does!
The results still look inaccurate. What's the best way to draw with
Because the search engine only looks at the strongest visual features
it will tend to produce more obviously "accurate" results if
you draw the most graphic images. Try keeping your images very simple
and direct, drawing only the most significant visual structure or skeleton
and skipping the details. So if you are drawing a face, concentrate on
the darkest features such as the eyes, mouth, hair and their relative
positions. Sometimes simple patterns can retrieve interesting matches.
If you draw a circle in the middle of the canvas it will treat it as a
completely image to if you draw it at the top left. And remember that
if an image of a face is half in shadow then it is that shadow silhouette
that will become the most likely feature for it to match on - it won't
recognise eyes themselves.
Where do the image results come from?
The images it retrieves are meant to reflect the scope of a commercial
internet search engine. But because I can't afford to store the millions
of images that a service like Google or Alta Vista can provide, the database
that Figga searches typically contains a few tens of thousands of images.
These have been collected at random from the internet and are periodically
updated. So it is a bit like a shifting "window" of all the
images on the internet. So if you submit the same image a few weeks later
you may get different results. The larger I make this database the more
accurate the results are likely to be - it can only return the nearest
matches to the images available in its current database.
How does it work?
Searching by visual appearance is known as CBIR (Content Based Image
Retrieval) and is still an emergent technology (that means it doesn't
work perfectly). It is an enormously more complex process than textual
matching because images are formed of continuous surfaces of features
and can vary in all the two dimensional aspects of size, orientation,
etc. They are not composed of a standard and limited set of signs. Figga
uses a piece of open source software called "imgSeek" for its
basic image recognition engine. imgSeek uses a technique called "multi-resolution
wavelet decomposition" to break images down into bits it can search
and compare. It is a bit like a compression algorithm like JPEG.
Who are you?
Richard Wright has worked as a media artist for over twenty years,
specialising in digital imagery, history and parallels between the Baroque
and digital culture. For more info see www.futurenatural.net.
Concept and design: Richard Wright
Technical Assistance: Tony Shaper
Hosted by: Furtherfield.org
Funded by a grant from Arts Council England 2006