REDMOND, Wash., Oct. 2, 2000 — Over time, most families will accumulate a large collection of photos, and many of those shots will probably end up in a shoebox on a closet shelf.
Today’s families use digital cameras to snap more than 1,000 photos each year, yet many people simply dump the shots into one directory, analogous to the “shoebox storage” method. Trying to find a photo in this directory is difficult — typically, shots are named something utterly nonintuitive, such as P000006.JPG.
At best, people go through a laborious process of setting up separate folders and trying to logically group their many shots.
But help is on the way. John Platt, a member of the Microsoft Research (MSR) Signal Processing group, has devised AutoAlbum, a new user-interface technology that automatically compresses a collection of digital images into a series of albums.
AutoAlbum displays a representative thumbnail image for each album. These thumbnails are shown as a contact sheet. When users find a likely album, they click on the representative thumbnail. AutoAlbum then displays a thumbnail “contact sheet” of the album’s contents. By clicking on any part of this second layer of thumbnails, a user can call up a full-size image.
AutoAlbum uses two techniques to group, or cluster, similar photos. First, the clustering algorithm looks at the creation time of the files, or the file names (which are usually generated sequentially when images are transferred to the PC), and according to the time they were taken, or the order in which they were taken, AutoAlbum then groups the photos into categories based on this information.
The second technique involves analyzing the color content of the images through a quick pixel analysis. Using a technique called Best First Model Imaging, the algorithm evaluates the order in which the photos were taken, looking for the two most similar adjacent photographs, groups them together into one album, and repeats that process.
“What the full algorithm does is to first cluster by time, looking for gaps in time,”
“It then tries to break photos up by similar color content.”
Platt used AutoAlbum technology to organize and condense about 250 of his own photos into about 23 thumbnails. A demonstration of the technology at http://research.microsoft.com/~jplatt/autoalbum/ex.html shows how AutoAlbum grouped mountain hiking photos separately from other shots of his furniture and of a party scene at a friend’s home.
The algorithm that powers AutoAlbum
“looks for clusters that are similar both in time and in color,”
“The reason why my party shots got separated from pictures of my furniture, for example, is because the two sets of photos were taken at separate times.”
The processing is speedy, too. Once the color-content analysis is completed, AutoAlbum can cluster 400 images in approximately three-tenths of a second.
Although Platt’s technology could be incorporated into Microsoft products now, this is unlikely to happen until next year. By then, AutoAlbum will be given an updated user interface, computer vision, and machine-learning techniques that are currently being developed by the Multimedia Database Working Group, a consortium of MSR researchers from Redmond; Cambridge, England; the Bay Area Research Center; and Beijing.
“AutoAlbum is the first step in a panoply of technologies that will, hopefully, improve a user’s experiences with images,”
For more information on the Signal Processing Group, please visit http://www.research.microsoft.com/signal/ .