Wednesday, August 25, 2010

Phone(y) Detectives!


Too many Apps are being developed that will turn people into ametuer detectives.

A new mobile application called Recognizr can identify a person’s face via your phone camera and deliver not only profile information about that person but also show you their latest status updates.

Swedish computer vision specialists Polar Rose combined forces with interface designers TAT (The Astonishing Tribe) to create the Recognizr as a prototype application for Android phones to show off Polar Rose’s mobile face recognition library. Polar Rose’s software recognizes individuals, while TAT’s interface uses augmented reality to show profile information from sites like Facebook, YouTube and LinkedIn and the latest status updates from the recognized person.

Recognizr uses FaceLib, a mobile face recognition library from Polar Rose, which is available for Android and iPhone. FaceLib can recognize faces in photo or video but, in common with other facial recognition products, is more accurate for photos. Recognizr also uses Polar Rose’s server-side solution FaceCloud because you can’t store profiles of all potential matches in the phone — although recognizing people who are already in the phone’s address book can be handled locally on the device.

In a presentation at the Hot Chips conference at Stanford University, David Petrou, a project lead of Google Goggles, described the future of Goggles.

Google Goggles, as Petrou reminded the audience, includes the ability for the mobile Android application to take a picture of the object ad send it back to Google's cloud services. Petrou demonstrated the app with a picture of a beer can with an Android smart phone, which identified the can as a can of Boddington's. A results page showed an icon of the result, with some results from the Web. He also showed a videocassette of the movie Breakin', and Goggles identified it correctly.

The basic design principle of Goggles, Petrou said, was that it has to be universal: queries can't be processed within a single finite context, such as a bottle of wine. Petrou showed off a book that contained an image of a manual transmission linkage: Goggle returned both a link to the book on a shopping site, but also linked to a search on manual transmissions.

Goggles returns a specific result about a third of the time, Petrou said, and the internal CONGAS recognition engine matches images to aa database of about 150,000 landmarks by finding "interest points" within an image. New photos compared to the database can be correctly identified about 50 to 60 percent of the time, with a false positive rate of about one in 10,000, he said.

Goggles' strong suits? Packaged goods, such as movies. But with generic objects, such as an image of a red car, Goggles still struggles.

Goggles can also work with bar codes. A recent addition has been the inclusion of machine translation, which can recognize text and translate it on the fly.

Unfortunately, Goggles has to work as a client application, as Google needs as much of a fine-grained control of the camera as possible, such as the white balance.

Petrou said that Google was considering opening up Goggles to third-party applications, so that a stamp collector could upload an image of a stamp with annotation describing what it is. An open API may also be released, so that a picture could be taken of a foreign currency, and an app could be opened to automatically convert that bill's value into dollars.

Google also plans to fuse the camera with Goggles, so that augmented reality may be the future of visual search, Petrou said. "We'll use it where it's the right user interface," he said.

Goggles does have the capability to recognize faces, although that functionality hasn't been implemented in the app as yet. That might change as more and more people begin uploading data to the Web: if 17 different images of your face appeared on the Web, a picture take of you with Google Goggles would rank "you" in the top ten results about half the time. If there were 50 results, your face would be ranked in the top 5 results abour half the time, he said

Facial Recognition along with Geo-tagging is also available in Picasa web albums. You can already find who is connected to someone in facebook when you search for the people. Once all these information would start working together then you can identify the person where they have been on holidays (based on Geo-tagging of photos) and when they have gone there. Who else in their friends/family had gone with them, etc.

In fact FBI/CIA/MI5 may make their own app to "report potential terrorists" where if you see any suspicious person you can open one of their apps and click the photo of some person and the secret service will quickly check if this person is potentially a problem.

Of course we are not discussing the privacy concerns yet but looks like Scary future to me.

No comments: