全局结构引导的人造物体参数化基元检测

Global Structure Guided Parametric Primitive Detection of Man-Made Objects

点云是表达三维数据的常见形式,点云数据提取出的几何基元能够帮助人们快速地理解并处理场景信息,也方便后续其他任务的开展.为了更好地利用人造物体中普遍存在的全局结构关系,增强基元检测过程中全局结构的正向引导,提出参数化基元检测网络——RelationNet,包括2个子模块.首先,为了更好地编码三维点与其所在基元的结构关系,通过空间偏移预测模块预测三维点所在基元中心的偏移向量,提升点对其所在基元的位置感知能力,为后续分割任务提供更多的特征依据;其次,人造物体的基元与基元之间常常具有如平行、垂直、轴对齐等结构关系,为了更好地利用这些关系实现对几何基元检测结果的改进,还包含全局结构关系提取模块,利用基元拟合后获得的参数判断各个基元之间的结构关系,并通过设置相应的损失函数对提取到的结果进行引导监督.在大型ABC数据集与基元监督拟合(SPFN),ParseNet等主流算法进行对比的实验结果表明,RelationNet在基元分割和基元分类任务上的MIoU分别达到85.32%和90.10%,与当前先进方法相比有明显的效果提升.

Point cloud is a common form of expressing 3D data.Extracting geometric primitives from point clouds can help people quickly understand and process scene information,and also facilitate other subsequent tasks.In order to better utilize the global structural relationship prevalent in man-made objects,this paper propose RelationNet to enhance its positive guidance in the process of primitive detection.RelationNet consists of two sub-modules.First,in order to better encode the structural relationship between 3D points and their primitives,RelationNet includes a spatial offset prediction module,which is used to predict the offset vector of 3D points to the center of the primitives where they are located.The location awareness of primitives also provides more valid features for subsequent segmentation tasks.Second,the primitives of man-made objects often have structural relationships,such as parallel,vertical,and axis alignment.In order to use these relationships to improve the detection results of geometric primitives,RelationNet also includes a global structural relationship extraction module,which uses the parameters obtained after primitive fitting to determine the structural relationship between each primitive,and sets the corresponding loss function to supervise the extracted results.This method is tested on the large scale ABC dataset and compared with mainstream methods,such as SPFN(supervised primitive fitting network)and ParseNet.The experimental results show that the accuracy of RelationNet in primitive segmentation and primitive classification reaches MIoU of 85.32%and 90.10%,respectively,which significantly outperforms state-of-the-art methods.