PureGin.org

Vancouver based Ludicorp's web photo-sharing site Flickr has been getting all kinds of (well-deserved, IMHO) attention and accolades in the popular media lately (The Manchester Guardian, Wired).

I've been quite interested in the possibility of using the context for images that Flickr creates to train image analysis software. Flickr has access to millions of images, with a fair amount of metadata, in explicit forms (tags provided by users) but also in implicit forms (images are grouped into sets according to owner; owners are related via the acquaintence network; viewers' interactions with images are to some extent derivable from session tracking). Some of this metadata is of dubious quality, but some is doubtless of very high quality indeed. This seems like a perfect data collection for people or organizations interested in training certain kinds of image analysis or image-mining software. It might be fun, for example, to work on 'auto-tagging' heuristics. Can we take an un-tagged image, and automatically apply descriptive tags (say from the 100 or 1000 most common tags), together with some kind of statistical measure of confidence? Can we automatically improve on the classification by seeing how users respond to the presentation of such automated metadata, say in search results? For example, say users searching for 'cat' were presented with an image of a flower; we could try to use the distribution of click-throughs to see which tag associations are anomolous.