Like I care about shoes...

Picasa enables me to find my photos on my own PC, but a lot of the power in that search comes about because of the good metadata associated with the images - I carefully file them in directories where the name means something, and by virtue of the date being stored with images I can often narrow the search down by knowing around when I took a photo (or more likely, whether it was before or after another photo).

On the web however, image searching is a different matter.  Metadata is much harder to find - images prepared for the web often have dates and EXIF metadata stripped off or altered, and images are not filed in neatly named categories.  Most search engines therefore rely on just approximate clues about the image - the filename, link text, other on page clues.  The result is that image search is a very poor relation to text search in many respects.

It's worth noting that there are a number of motivations for doing an image search:

• to find an image of a particular item - this may be the obvious one, but its actually not one that drives most of my searches
• to return to a website that I remember was illustrated by a particular image
• as a sort of visual snippet to help decide which sites are worth visiting (on the basis that if a site goes to the trouble of sourcing good images, the rest of its content may be of a similar standard)

Thus around a year ago when Riya launched, I was hoping for a site that would push the boundaries of image search.  Their particular hook was that they did textual and facial recognition.  The facial recognition got the big press, since it was the sexier technology, but was beset with problems - it didn't work well enough to live up to the hype, and had it done so, there are all sorts of privacy concerns, with shades of Big Brother being able to find lots of images of an individual automatically.

I was therefore more interested in the textual recognition that they offered.  However, if possible, this was even worse than the facial recognition.  The recognition rate was quite low - many images showing text had nothing in them recognized, and even when text was recognized, it was normally as single words, rather than as a block of text.  (My understanding from Tara Hunt is that this was not a technical issue, rather that the licence to the technology that Riya obtained to do this would not permit them to do continuous text recognition).  There are a few images that the text recognition has tagged - for example search for "exit", but in general so little text from images was recognized that you probably forget the site ever offered this feature.

Riya realized they had problems, (their CEO writes a very open blog on the matter), and decided to change their emphasis, to become a real visual search engine.  This sounded great, and I eagerly awaited the arrival of Riya 2.0 as they considered it.

And today it arrived - and I don't like it.

Rather than perfecting their technology for a wide ranging search field, they have instead produced a shopping service - you can use it to search for products that look like each other - provided that those products are shoes, watches, handbags or jewelry.  I'm certainly not in their target market - 20 to 30 year old women, so perhaps I'm missing the point, but these seem a strange set of goods to handle.  Shoes are one of the last things I would buy on-line - unlike most clothes they need to fit pretty exactly, so generally need to be tried on in person.  Watches and jewelry are high value items - and not everyday purchases, so not an obvious online purchase either.  And as a man, I'm never likely to understands handbags - so again the exact logic behind offering this product escapes me.  Perhaps the reasons for these choices lies with the celebrity pictures to be found on the site - where the pictures show watches and shoes highlighted, and linked to searches.

All in all, with the widespread coverage of like.com's launch, it does look as if I'm not alone in thinking they could have done more with this technology.  Here's hoping that Picasa is able to do more with the related object recognition technology they recently got via Neven Vision.

Good further coverage of Like is to be found at: Robert Scoble, Thomas Hawk, ZDNet, TechCrunch, Greg Linden, GigaOM, Don Dodge.

It's worth noting that visual search for shoes is not unique to Like - Chez Imelda also offers this, powered by technology from Pixsta.