Groups of mathematically strong and weak second-, fourth- and sixth-graders were individually confronted with numerosities smaller and larger than 100 embedded in one-, two- or three-dimensional realistic contexts. While one third of these contexts were totally unstructured (e.g., an irregular piece of land jumbled up with 72 cars), another third had a clear structure (e.g., a 16x4 rectangular parking lot completely filled with 64 cars), and a last third had a "semi-structure" (e.g., the same 16x4 parking lot but with a number of cars missing). Besides analyzing the effects of different task and subject variables on pupils' accuracy and response-time data, the study involved an analysis of their estimation strategies, with an emphasis on multiplicative strategies that profited by some of the tasks' geometrical (semi-)structure. It was found that many children actually made use of such strategies, that using these strategies did however not always led to accurate estimations, and that their frequency and efficiency increased with age. (Contains 4 figures, 5 tables, and 1 note.) [Abstract is presented in both English and French.]