construction3d videonetworkcalibrationsilhouette   technology
In order to produce a Euclidean 3D model, all cameras have to be fully calibrated with respect to a single world coordinate system. Classical methods use special calibration patterns with known 3D coordinates. For reasons of accuracy, they should fill the whole working volume, which disqualifies them for the blue-c. Additionally, such patterns would not be visible to all cameras, thereby calling for a chained, sequential calibration in which errors tend to accumulate.

Instead, we propose a fully automatic procedure based on self-calibration. A person waves a standard laser pointer in the darkened blue-c, thus filling the working volume with virtual points. A small piece of transparent plastic is attached to the pointer for better visibility. Geometric constraints are used to clean the set of points in the different views: false points are removed and missing points are added. Their projective structure is computed by rank-factorization and refined through bundle adjustment.

The Euclidean stratification is computed by enforcing orthogonality of camera rows and columns and is initialized by putting the principal points into the image centers. This self-calibration procedure yield all parameters of the pinhole camera model. This linear model initializes a refinement procedure which yields parameters of the radial distortion. Reprojection error of the complete projection model significantly below 1/2 pixel is achieved even for very wide angle lenses with huge radial distortion.

[Tomas Svoboda]

standard laser pointer graphical output validation
standard laser pointer graphical output validation