<p>Point cloud has been accepted as an efficient data format for 3D object recognition. Despite its successful usage in common object recognition, in the industrial domain, where mechanical components are the main recognition targets, voxels or multi-view images are still the more commonly used 3D representations since they result in better recognition accuracy. Yet, compared with point clouds, voxels or images also lead to heavy computation and thus slow down the recognition speed. In this work, we argue to use point clouds to achieve real-time mechanical recognition, meanwhile attempting to improve the recognition accuracy by adopting novel learning architectures. We start by analyzing the unique 3D properties of mechanical components that may affect the accuracy of point cloud recognition. We find that mechanical components are usually unaligned and can have millimeter-level inter-class differences. Both properties are unfriendly for point cloud representation, which is sensitive to coordinate rotation and may be too sparse to capture the millimeter-level differences. To this end, we present the Multi-View Point Cloud Transformer which comprises two key modules to alleviate the above problems. First, a Multi-View System, that converts the rotation-sensitive point coordinates into rotation-invariant multi-view observations to deal with the unaligned inputs. Second, a Multi-View Transformer, which reasons about correlations across multi-view observations via a novel decoupled local attention, capturing the small inter-class differences for recognition. Experiments confirm that the two modules effectively collaborate to recognize both mechanical components and common objects, achieving faster speed and better accuracy than most existing methods.</p>