The technique views even flying butterflies

Jun 18, 2009 19:01 GMT  ·  By
Boston College Assistant Professor of Computer Science Hao Jiang and colleague Stella Yu report they've developed a novel solution of linear algorithms to streamline computer visualization
   Boston College Assistant Professor of Computer Science Hao Jiang and colleague Stella Yu report they've developed a novel solution of linear algorithms to streamline computer visualization

Drawing their inspiration from the anatomy and the physiology of the human eye, researchers at the Boston College have devised a new viewing technique that allows computers to see fleeting images, such as butterflies flying and fish swimming very fast. Its accuracy is double, and its speed is ten times that of any other existing technology of the same sort.

The researchers involved in the study say that fields of research such as action and object recognition, surveillance, wide-base stereo microscopy and three-dimensional shape reconstruction could benefit a great deal from the innovation, bringing about a new wave of fresh and powerful devices. Details of the new technology will be presented later this year, at the upcoming annual IEEE meeting on computer vision. In charge of the new research were computer scientists Hao Jiang and Stella X. Yu, both from the Boston College.

In earlier versions of similar systems, a video camera mounted on the computer would capture the image, and the computer would then browse through millions of pictures, looking for a possible match. This task was made more difficult by moving objects, where shifting orientations and angles continuously modified the search parameters that the search software used, forcing it to start over.

This task was very time- and resource-consuming, and its results were questionable, to say the least. In the new method, the research team was able to devise a novel solution of linear algorithms, which is able to make the computer's work a lot easier.

“When the human eye searches for an object it looks globally for the rough location, size and orientation of the object. Then it zeros in on the details. Our method behaves in a similar fashion, using a linear approximation to explore the search space globally and quickly; then it works to identify the moving object by frequently updating trust search regions,” Jiang, who is a BC assistant professor of computer science, explains.

Unlike other viewing methods, which only have a 50 percent success rate – while consuming enormous amounts of resources during a session –, the new technology is able to pick up objects ranging from teddy bears to moving fish with a 95 percent accuracy, at only a fraction of the hardware complexity of its predecessors. It relies on mathematical formulas to identify an object, a process that functions a lot faster and more efficiently than comparing data-bank pictures with those obtained by an outside camera.