Carnegie boffins develop software that makes pictures really 'perfect'

Report from the Asian News International brought to you by the Hindustan Times

London, Aug. 12 -- Researchers at the Carnegie Mellon University in Pennsylvania, have developed two computer programmes that literally makes pictures 'perfect'.

The first program - dubbed the Scene Completion software - allows photographers to remove unwanted people or objects from their digital snaps and replace them with perfectly matched alternatives.

The programme takes into account key measurements, such as camera angles and lighting conditions, and looks through more than two million photographs on Flickr, an image - sharing website, selecting a range of pictures that can be blended on to the original photograph.

As such, a stray thumb ruining a beach shot can be replaced with a bucket and a spade, or a building blocking a coastline view can be removed in favour of miles of glistening sea, they said.

Also to make sure nothing absurd is blended into the picture, the software searches for objects already present in the original and then fixes in the best match, they added.

The second approach uses Internet-based image libraries to create a clip-art of objects that, once inserted into a photograph, looks convincing.

To generate the clip art, the team first drew on a series of online image libraries offering cutouts of different objects. The software then estimated the height of the camera used to take the picture, working out the distance between objects in the snap.

Looking at the distribution of colour shades within three broad regions: ground, vertical planes and sky, the software then worked out the light sources in the scene.

With knowledge about the position, pitch and height of the camera and light sources, the software finally looked for images in the clip art database taken from similar positions.

The scientists say the results are so good that during a series of trials, the altered snaps passed as originals 70 per cent of the time.

They said although editing programs in vogue, such as Photoshop allowed snaps to be tweaked and elements erased or replaced, they were often time-consuming and the results unconvincing.

"Matching an object with the original photo and placing that object within the 3D landscape of the photo is a highly complex problem. But with our approach, we can hide the complexity from the user and make the process simple and intuitive," said Jean-Francois Lalonde, who headed the research.

He added the success of their software depended on the number of photos available to scan.

"We saw a dramatic improvement when we moved from a database of 10,000 images to two million images. And that is just a tiny fraction of the hundreds of millions of images already available on sites like Flickr," said researcher James Hays.

"We have tons of photos from which to choose," the Daily Mail quoted him as saying.

Published by HT Media Ltd. with permission from Asian News International.