Precision in 3-D Points Reconstructed From Stereo

We characterize the precision of a 3-D reconstruction from stereo: we derive confidence intervals for the components (X,Y,Z) of the reconstructed 3-D points. The precision assessment can be used in data rejection, data reduction, and data fusion of the 3-D points. Also, based on the confidence intervals a bad/failing stereo camera pair can be detected, and discarded from a polynocular stereo system. Experimentally, we have evaluated the performance of the confidence intervals for Z in terms of empirical capture frequencies vs. theoretical probability of capture for a test, ground truth, scene. We have tested the interval estimation procedure on more complex scenes (for example, human faces), but since we do not have ground truth models, we have evaluated the performance in such cases only quantitatively. Currently we are developing "ground truth" models for more complex (such as general indoor) scenes, and will evaluate quantitatively the performance of the confidence intervals for the depth of the reconstructed points in the "automatic" rejection of 3-D points which have high degree of uncertainty.