The “never solved” problem of parametric estimation in computer vision is presented in this paper. This problem is particularly difficult because of the non-linearity of measurement equations, the parameter constraints, and the errors introduced by acquisition and quantization. These difficulties call for a robust estimation in the presence of outliers as well.

First the estimation problem is defined as an optimization problem. Then, the authors describe how to consider the problem as a projection problem, for which efficient and well-established solutions have been reported in the literature.

In section 3, the problem of outlier rejection is addressed by means of a classical randomized estimation method, and a measure of the relevance of this estimation is proposed. In addition, the authors propose the use of a hierarchy of models to better estimate parameters, and describe how to implement the robust estimation in this case.

Eventually, the software architecture of the estimation module is discussed, and the results of some experiments are reported for some examples of the problem of “fundamental” matrix estimation.