Building a generic and performant vision system is still an open research problem. Actually, real-world vision systems need to face the challenge of dimensionality and ambiguity of data. To tackle this problem, we introduce a dynamic computational attention model. This latter selects the most salient scene information while being able to adapt its behaviour to the needs of a generic vision system. We focus validation on plusibility, to check this property we compare through three measures results obtained by several algorithms and ground thruth.