Dear editor, Three-dimensional (3D) object detection is a fundamental computer vision issue and demonstrates promising potential in intelligent surveillance, robot grasping, and autonomous driving. However, 3D object detection remains a challenging problem because it must predict many object details in 3D space, ie, depth, shape, and orientation. Current 3D object detectors involve onestage and two-stage frameworks. The two-stage framework first uses a region proposal network (RPN) that produces non-uniform region proposals and converts their region-wise features into an identical form. Then, the identical features are sent to another network to filter and refine the initial proposals. The one-stage framework adopts a hypothetical region proposal scheme that imposes hypothetical non-uniform proposals to certain activations of the network and thus removes the region of interest (ROI) pooling block. One-stage detectors ignore the statistic shape priors of the objects involved. Obviously, shapespecific proposals would aggregate more discriminative features and thus enhance detector performance, which is virtually the same as the top-down attention scheme [1] discovered in the human vision field.