Riya unleashing API to integrate its visual search into apps

Riya CEO Munjal Shah has been blogging about his company's shift in strategy, moving from a focus on face recognition to visual Web search (see also my post from May 16 and recent posts from Mike Arrington and Matt Marshall, among others). "The idea is to do similarity search globally, both in faces and sections of photos," Shah said. "If there is a car in a selected section, it will look for the blobs of pixels with the same shapes."

On Tuesday Riya is publishing a REST (REpresentational State Transfer) API that opens up the core facets of its software to application developers. "The new API opens all the functionally of Riya to application developers on a non-commercial basis, with commercial use pending request and review. The functions range include face detection, face recognition, text detection, text recognition, search, tagging and even building training sets," Shah said.

A basic use of API could enable an application to use Riya's face detection technology to find faces, autocrop the images and then send them to another application. A more complex application could take images of rooms, apply some rules about matching colors and shapes, and then look for images that are similar to the request, with all the image parsing and processing done by Riya's servers, Shah explained. "APIs for most companies don't yield compelling commercial applications," Shah said."It takes some nurturing and collaborative effort to build something great, but a lot of photo sites can add Riya's API as an enhancement."

Riya has started crawling Web sites for photos and calculating visual signatures, such as color histograms and dominant shapes, and storing them as mathematical vectors. The new search engine, Riya 2.0, will be open to the public in about three weeks, he said. Riya also plans to have a browser extension that initiates a similarity search via a mouse click.

Similarity presents some problems in generating useful results. "We don’t know how well similarity search will work in each category, such as real etate, people, or products. It may work equally well from an algorithmic perspective, but dating, for example, is a more exacting kind of  application. With objects, users may be more intolerant of incorrect results," Shah said. "For similarity, there is no right answer--many people can look similar to you. We'll see if it drives a different user behavior." Riya will have a feature that allows users to refine similarity searches by clicking "more like this" and "less like this" controls on each image.

Clearly, the revenue opportunities in a visual search engine are more extensive than face recognition of photos uploaded on Riya, with the potential to tap into lead generation and advertising fees that fuel image search sites like Yahoo and Google. For example, a user selects an image of a product, Riya 2.0 finds similar objects and generates revenue via clicks and ads on search result pages. It also turns out the searchers outnumber the uploaders on the current Riya service 20 to 1, Shah said. Riya could also leverage tags and other metadata to improve search similarity results.

Shah speaks of photos as the "thousand word query"--a picture worth a thousand words. He admits that so far, Riya has only uncovered only a handful of the metaphorical thousand words required to make visual search and image recognition really smart. But, it is just the beginning, and so far Riya is leading the way...