基于子空间及多示例多标签的三维数据分割与标注

Effective and efficient 3D data analysis, management, integration and reusing heavily depend on accurate and high quality 3D data semantic segmentation and labeling. Using the subspace analysis and multi-instance multi-label learning as the core technologies, this project will mainly study on the (dynamic) model set hierarchical semantic segmentation and weak supervise indoor scene semantic segmentation and labeling. The details are in the following:.(1) For the model set hierarchical semantic segmentation, we firstly discuss the enhancement of the presentation for the local feature and structural characteristic, based on the network technology. In this situation, the distribution of feature space is described by subspace. Then the geometry and corresponding of models are used to constrain the process of subspace extraction. Finally, the model set hierarchical semantic segmentation is achieved by this process..(2) For the dynamic model set hierarchical semantic segmentation, the hierarchical segmentation of new added model is completed by the technology of subspace recognition. The identification error is used to construct the core evaluation criteria. Moreover, the validity of segmentation and the compatibility between different levels are used to improve the globality of the hierarchical semantic segmentation..(3) For the scene labeling, we intend to combine the multi-instance multi-label learning with the especial prior knowledge that the function of labels and the way of interaction between labels should be consistent. Based on this combination, we will design an algorithm for the weak supervise indoor scene semantic segmentation and labeling. Benefited from this algorithm, a quality label classifier can be obtained, without the label of each pixel in the training data.

三维数据语义分割与标注是大规模三维数据分析、管理和重用的重要基础.本项目拟以子空间分析和多示例多标签方法为核心,对(动态)三维模型集的多层次语义分割和弱监督室内场景语义分割与标注展开深入研究..(1)针对三维模型集的多层次语义分割,基于网络技术,探讨增强模型局部特征和结构特征表达,在此基础上,利用子空间对特征分布进行刻画,并利用模型的几何结构、对应关系对子空间的提取过程进行约束, 实现模型集的多层次语义分割..(2)在动态模型集的多层次语义分割方面,从子空间识别技术出发,实现新模型的增量分割.利用识别误差构造核心评价准则,结合分割有效性及层次兼容性,实现多层次的总体约束..(3)针对三维场景标注问题,以多示例多标签技术为核心,结合标签间的空间构成、力学作用及功能一致性先验,构建弱监督室内场景语义分割与标注算法,在避免对训练数据进行繁琐的像素级别标注的同时学习获得高质量的标签分类器.

三年来，项目组主要围绕三维及更高维数据的语义及结构分割问题展开研究，分别提出了基于子空间结构的点云局部结构多层次分割算法、基于鲁棒非负局部坐标分解的弱监督分割方法、基于增量低秩表示的在线分割算法、基于特征插值卷积的三维模型/场景的语义分割算法、基于子空间学习的鲁棒低秩表示分割算法等一系列分割算法。本项目执行期间，项目组成员团结协作，取得了一系列有意义的研究成果，构造了首个三维点云数据局部结构分析数据库，为三维点云数据局部结构分析提供了学习和比较的依据，在IEEE Transactions on Visualization and Computer Graphics (TVCG), IEEE Transactions on Image Processing (TIP), Pattern Recognition (PG), Neural Networks等国内外刊物上发表论文13篇，其中SCI 检索11篇，北大核心期刊《计算机辅助设计与图形学学报》2篇。参加国内学术会议2次，国内外线上研讨多次。项目共计培养硕士研究生1名，在读硕士研究生6名。