The only assumption which is strictly needed by the method is that pixels are rectangular (see for a proof [144,131]). In practice however it is interesting to make more assumptions. In many cases pixels are square and the principal point is located close to the center of the image. Our systems first uses a linear method to obtain an approximate calibration. This calibration is then refined through a non-linear optimization step in a second phase.
In Figure 6.3 the retrieved structure and motion is shown before (top) and after (bottom) self-calibration. Note that metric properties such as orthogonality and parallelism can be observed after self-calibration.
