DARPA , NIMA , Science & Technology Foundation and INCO
[Ja94] C.0. Jaynes, F. Stolle and R. Collins. “Task driven perceptual organization for extraction of rooftop polygons,” in proc. Image Understanding Workshop, 1994.
[Ma99] M. Marengoni, C. O. Jaynes, A. R. Hanson and E. M. Riseman. “Ascender II, a visual framework for 3D reconstruction,” in ICVS, pp. 469-488, 1999.
[Gr02] E. Grossmann. “Maximum likelihood 3D reconstruction from one or more uncalibrated views under geometric constraints,” Ph.D. thesis, 2002.
[Gr04] E. Grossmann and J. Santos-Victor. “Least-squares 3D reconstruction from one or more views and geometric clues,” submitted to CVIU, 2004.
In photogrammetric 3D reconstruction, the image projections of 3D points of interest are identified in images and these 2D points are used to obtain the 3D reconstruction. Photogrammetric reconstruction exploits the fact that the optical center, observed image point and corresponding 3D point are aligned, i.e. the camera is a perspective projection device. Starting from the 19th century, this alignement constraint was used to compute the 3D points from two or more images produced by a camera whose focal length and other characteristics. However everyday cameras are uncalibrated, i.e. these characteristics are unknown, so that the images they produce cannot be used directly by photogrammetric methods.
More recently, it has been shown possible [ Fa92 , Ha92 ] to estimate the calibration parameters of cameras from identified image points alone, a process called self-calibration. These findings allow to obtain reconstructions from the more commonly available data.
It was shown [ Bo93 , De96 ] that 3D reconstruction is made easier when one exploits some geometric properties of the scene. Elements common to man-made environments -planes, right angles, repetitive and symmetric patterns- often allow to obtain a reconstruction from a single image, something that could not be achieved previously [ Sp92 , Sh98 , Cr99 , St99 ].
We show [ Gr02 , Gr04 ] how to obtain reconstructions in more general conditions still, by exploiting more geometric properties and allowing single- and multi-view datasets to be treated alike. Also, we developed a parameterization of 3D points subject to geometric constraints that uses only common linear algebra software tools and is suitable for integration in numerical optimization algorithm. The resulting reconstruction is a least-squares estimate whose precision can be assessed.
Finally, an important goal in photogrammetric 3D recontruction is the ability to obtain a reconstruction of extended urban areas fully automatically, while correctly estimating the architecture of buildings. This is an important goal because interactive reconstruction methods are very labor-intensive, while traditional fully automatic photogrammetry methods do not correctly model urban scenes.