基于对称形状生成的三维目标检测网络

3D object detection network based on symmetric shape generation

基于点云的三维目标检测在机器人、自动驾驶等领域起着至关重要的作用,激光点云能为场景理解提供精确的几何信息。然而,由于点云的稀疏性和物体间的遮挡关系,激光点云通常只能描述物体的部分形状,导致目标结构信息不完整,从而降低检测精度。针对这个问题,提出基于对称形状生成的三维目标检测网络(SSG-RCNN),一种双阶段目标检测器。考虑到感兴趣目标形状的对称性,SSG-RCNN在一阶段生成候选框的同时为每个前景点预测镜像对称点,从而恢复目标的整体形状。二阶段中,使用自注意力池化层从原始点和对称点中聚合特征用于候选框修正,完成三维框预测。KITTI数据集上的实验表明SSG-RCNN取得了卓越的检测性能,在测试集上对困难目标的检测精度达到77.64%,高于所有对比方法。

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 two-stage 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 SSG-RCNN 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.