In this paper we address the problem of the choice of visual primitives structure in multi- images analysis systems. We consider that this problem has a great consequences over the robustness of such systems, over their using in different vision tasks and over their computation time. Usually, methods used in image analysis systems (pattern recognition, 3D reconstruction, ...) tend to match visual primitives like pixels, regions or edges to recognize an object, to search candidate couples, and more generally to analyze multiple images. In this work we clearly demonstrate that the structure of such primitives is extremely important to obtain robust systems. We don't want to compare between different primitives, but between different structures that one of them can be represented. The primitive we have chosen, in this discussion, is the edge of objects. Finally, complete quantitative experimental results are shown with various indoor and outdoor real world senses. The system has been tested on both stereo pairs and images sequence, which are two different applications.