面向点云数据的三维风格迁移研究

Three-dimensional point clouds can accurately and stereoscopically describe the real world scene, which have obvious data advantages than two-dimensional images. However, point cloud data obtained though existing laser scanning systems often exhibit inconsistencies in local details, resulting in unsatisfactory performance in realism. This research starts from non-photorealistic rendering, and focuses on three aspects: the multi-scale supervoxels based point clouds description, the three-dimensional stability of style features in artistic images, and the visual coherence of multi-view style images. The research explores the theory and methods for transferring the styles from two-dimensional art images to three-dimensional point clouds. The major innovations include: (1) Researching on the subset selection based multi-scale supervoxels segmentation for point clouds description; (2) Researching on the scientific problem of three-dimensional stability style features decomposition; (3) Researching on a multi-view projection based three-dimensional style transfer method which combines two-dimensional and three dimensional style features. The output of this research will enhance the artistic and visual perception of the point clouds, and benefit the applications such as virtual reality, animation games, and multimedia.

三维点云可以精确、立体地描述并数字化再现真实世界场景，相比二维图像具有明显的数据优势。然而，通过现有的激光扫描系统获取的点云数据常常出现局部质量不一致的现象，导致在真实感的表现上不尽人意。本课题从非真实感出发，围绕多尺度超体素的点云描述、艺术图像风格特征的三维稳定性、以及多视角风格图像的视觉连贯性三个方面展开研究，探索从二维艺术图像到三维点云的风格迁移理论和方法。主要创新点包括：（1）研究基于稀疏子集选择的多尺度超体素分割的点云描述；（2）研究艺术图像风格特征的三维稳定性分解这一科学问题；（3）研究基于多视角投影的融合二维和三维风格特征的点云风格迁移方法。本课题的研究成果将直接推动点云数据在虚拟现实、动漫游戏、多媒体等领域的应用。

三维点云可以精确、立体地描述并数字化再现真实世界场景，相比二维图像具有明显的数据优势。然而，通过现有的激光扫描系统获取的点云数据常常出现局部质量不一致的现象，导致在真实感的表现上不尽人意。本课题从非真实感出发，围绕多尺度超体素的点云描述、艺术图像风格特征的三维稳定性、以及多视角风格图像的视觉连贯性三个方面展开研究，探索从二维艺术图像到三维点云的风格迁移理论和方法。项目完成了以下研究内容：（1）研究基于稀疏子集选择的多尺度超体素分割的点云描述；（2）研究艺术图像风格特征的三维稳定性分解这一科学问题；（3）研究基于多视角投影的融合二维和三维风格特征的点云风格迁移方法。研究成果直接推动了大规模三维点云处理领域的发展，在虚拟现实、多媒体、数字城市等方面具有重要意义。