3D object detection based on point cloud is essential in many applications, such as robotics, autonomous driving. LiDAR point clouds contain reliable geometric information for 3D scene understanding. However, due to sparsity and occlusion, point clouds depict only partial surfaces of objects, which severely degrades the detection performance. To handle this challenge, we propose a novel twostage detector based on symmetric shape generation ( SSG-RCNN). The shapes of 3D interested objects are roughly symmetric. In the first stage, SSG-RCNN predicts a symmetric point for each foreground point to complete objects shapes while generating 3D proposals. In the second stage, SSG-RCNN utilizes self-attention pooling module to aggregate proposal-wise features from raw points and symmetric points. Finally, proposal-wise features are used to refine 3D proposals. Extensive experiments on KITTI benchmark show that SSGRCNN has remarkable detection performance. Especially for hard difficulty level objects, SSG-RCNN achieves 77. 64% AP on the KITTI test set, which is better than previous state-of-the art methods.