thanks for the feedback! It is the same idea of using a limited annotation vocabulary, but used for the purpose of helping ameliorate poor text metadata in bulk search, rather than suggesting tags to users. Right now automatic annotations are deliberately not combined with text metadata, to demonstrate raw capabilities of the former. You can try how metadata search would perform on its own by using it as a separate search option. We give a short rundown of how the system works here and we even have a blog for everyone to leave comments.

I had a quick look at Behold – seems a similar idea in terms of classifying images using a limited vocabulary of annotation tags. Is that right?

I tried ‘beach’ as an example visual search, and was surprised that the first couple of results seem less relevant than later results. Is there any attempt to rank/order the images presented like in text search, or is it simply a set of images with annotations that match the keyword? I was wondering, for example, whether you take into account multiple forms of evidence for image relevance: e.g., your automatic annotations combined with meta-data around the image, plus the image file name or other similar properties.

Thanks for the link.

I was actually more impressed by Riya’s work in face recognition for photos than the “likeness-based” search, particularly given it is limited to very constrained domains of application right now.

More generally, I’m just not sure how useful image-input search is. A lot of the time, when I’m searching for an image, I don’t have a prototypical image to use as an input query. Even if I do, I’m not sure that you could adequately find similar images using image processing algorithms (I don’t just want something with similar colouring, contrast etc, but the same subject matter most of the time – and that’s *much* harder to detect).

But I guess, as you point out, it may have application in some limited domains. Perhaps even the fashion domain it is currently targeted at is one such domain – I don’t know.

They’ve gone a slightly different route, and rather than use tags to find images, they use an existing image.

The image processing algortihm seems good, but not good enough for general use. However, in the targeted verticals they are using it for … wow!