Bringing computer vision into focus

By W. Conard Holton At the IEEE Computer Society Conference on Computer Vision and Pattern Recognition this year, Microsoft Research (Redmond, WA, USA; research.microsoft.com) presented papers highlighting research in computer vision. These ranged from noise removal from digital images and techniques for building 3-D models to handwritten signature verification and image recognition and search algorithms.

Nov 12th, 2007

By W. Conard Holton

At the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Minneapolis, MN, USA; June 18--23; cvpr.cv.ri.cmu.edu) this year, Microsoft Research (Redmond, WA, USA; research.microsoft.com) presented papers highlighting research in computer vision. These ranged from noise removal from digital images and techniques for building 3-D models to handwritten signature verification and image recognition and search algorithms.

In their paper "Removal of image artifacts due to sensor dust," Changyin Zhou of Shanghai Fudan University (Shanghai, China; www.fudan.edu.cn) and Steve Lin, a researcher for Microsoft Research Asia, offered a technique to remove dust effects from digital photographs automatically. "Currently," Lin says, "these dust effects in photographs must be removed by hand using software. We provide an automatic tool for dealing with this problem."

The key idea," Lin explains, "is to model the physical formation of dust artifacts and use to obtain results superior to those of general artifact-removal techniques."

In his paper "Inferring temporal order of images from 3D structure," Sing Bing Kang, a senior researcher at Microsoft Research, describes a technique he has developed to order chronologically a large set of photographs by analyzing the presence of structures visible in the photographs. "Given a large collection of historical photographs of a city that spans decades or perhaps even a century," Kang says, "how can photographs automatically be sorted by time? We attacked the problem by reasoning about structures that are visible in photographs. This is made possible by assuming that each structure is unique and exists in a continuous block of time."

"Sorting photographs in time," Kang says, "allows time-varying 3-D models to be constructed from collections of historical photographs. The user can then navigate in space and time. "Imagine virtually turning back the clock and visualizing a city as it had appeared in the past. We imagine such a tool would be useful to historians and urban planners."


More in Boards & Software