We present a linear algorithm for self--calibration which, like the
Sturm/Triggs method for projective reconstruction, is exact in the
limit of small camera motions or sideways motions with rotations around the
optical axis. Unlike previous algorithms, our approach recovers the
available partial information about the internal parameters for ``critical''
motion sequences, where full recovery of the calibration is impossible.
Also, we present a linear algorithm for estimating the camera motion and 3D
structure given a projective reconstruction and known calibration, and we
extend our linear approaches to more accurate, iterative algorithms. All our
algorithms follow the stratified approach in which the Euclidean
motion, 3D structure, and calibration are computed from
a projective reconstruction.