Automatic urban object detection remains a challenge for city management. Existing approaches in remote sensing include the use of aerial images or LiDAR to map a scene. This is, for example, the case for patch-based detection methods. However, these methods do not fully exploit the 3D information given by a LiDAR acquisition because they are similar to depth map. 3D Deep-Learning methods are promising to tackle the issue of the urban objects detection inside a LiDAR cloud. In this paper, we present the results of several experiments on urban object classification with the PointNet network trained with public data and tested on our data-set. We show that such a methodology delivers encouraging results, and also identify the limits and the possible improvements.