In this paper, we propose a new approach for the classification of 3D objects inspired by the Time Series Shapelets of [Ye and Keogh, 2009]. The idea is to use discriminating sub-surfaces for the current classification in order to take into account the local nature of the relevant elements. This allows the user to have knowledge concerning the sub-parts that have been useful for determining the belonging of an object to a class, and to get a better classification rate than current state of the art methods. The results obtained confirm the advantage of the random selection of candidate characteristics for the pre-selection of attributes in supervised classification.