This paper proposes a novel approach to the retrieval of buildings' height from multi-angular high spatial resolution images. To achieve this task, we combined two main concepts: multilevel morphological attribute filters, used for the definition of the objects in the image, and geometric invariant moments exploited for the characterization of the spatial properties of the previously detected shapes. The main concept of this study relies on the spatial properties of very high resolution images acquired from different angles of view. In particular, if we model the urban environment as an ensemble of vertical and horizontal surfaces, we can assume that the shapes related to the horizontal surfaces (i.e. the top of the buildings) do not suffer any relevant spatial distortion if detected from two angles of view, while vertical surfaces present strong changes in shape. Starting from this assumption, once each shape in each angular images has been spatially characterized, it is possible to identify univocally the same horizontal surface (i.e. the roof of a building) in each angular image. Finally, the knowledge of the acquisition angles permits the retrieval of the buildings height using simple trigonometric calculations. In this paper the proposed approach has been successfully applied to a WorldView-2 (WV2) very high resolution dataset composed by five angular images.