This project page is just a placeholder for older browsers. To see the most recent developments regarding the project, please enable JavaScript and refresh this page.

We are interested in the use of advanced statistical methods to cope with the important problem of visual correspondence. When reconstructing a 3D model from a set of images, a crucial question is which visual features in the different images correspond to each other. We have shown that the problems of 3D modeling and of correspondence can in fact be solved simultaneously, when viewed within the proper probabilistic framework. Our research group is taking this problem of matching across views to the next level, matching across space and time, in a project called 4D Cities. The challenge is that of reconstructing a 3D model of a city as it evolves over time, from unlabeled historical photographs.

Frank Dellaert (PI)

People

Relevant Publications

• Atlanta World: An Expectation-Maximization Framework for Simultaneous Low-level Edge Grouping and Camera Calibration in Complex Man-Made Environments, CVPR 04
• Multiview Reconstruction of Piecewise Smooth Subdivision Curves with a Variable Number of Control Points, ECCV 04
• Spectral Partitioning for Structure From Motion, ICCV 03
• EM, MCMC, and Chain Flipping for Structure from Motion with  Unknown Correspondence , Machine Learning Journal, 2003
• Monte Carlo EM for Data-Association and its Applications in Computer Vision, Ph.D. Thesis, 2001
• Feature  Correspondence: A Markov Chain Monte Carlo Approach, NIPS 00
• Structure from Motion without Correspondence, CVPR 00