Error-sensitivity assessment of vision algorithms

A methodology is proposed for the computation of output-data errors from given input-data errors for a given computational precision for vision algorithms. The approach is based on replacement of the arithmetic types of the object-oriented language used for coding of the tested algorithm by structures describing the values and error models of the numbers represented. The method is useful both for the initial testing of any vision algorithm and for the definitive testing of algorithms for which explicit error models are not known. It is also shown how two error models, namely that of the min/max error propagation and that of local-function linearisation, can be used for assessing the errors introduced by the input data and by the hardware architecture adopted.