Processing large amount of high-resolution 3D data requires enormous computational resources. As a result, a suitable 3D data representation must be chosen, and the data must be simplified to a size that can be easily processed. The question is how can the data be simplified? Random point sampling is a common sampling strategy. However, it is sensitive to changes in density. We build a sampling module based on a hybrid model that combines point cloud and voxel data. To determine the relationship between points within each voxel, the module uses the magnitude of the point (the Euclidean distance between the point and the object’s center) along with angles between each point embedded within each voxel. By exploiting farthest point sampling (FPS) that begins with a point in the set and selects the farthest point from the points already selected iteratively, our method has the advantage of covering the whole point set within a given number of centroids and still maintains the key benefits of both point cloud and voxel to better characterize geometric details contains in a 3D shape. With further observation that the number of points in each cell differs, we use a point quantization method to ensure that each cell has the same number of points. This allows all voxels to have the same feature size vector, making it easier for 3D convolution kernels to extract object features. We demonstrate these benefits and make comparisons with solid baselines on ModelNet10, ModelNet40 and ShapeNetPart datasets, demonstrating that our method outperforms some deep learning approaches for shape classification and segmentation tasks.

处理大量高分辨率 3D 数据需要巨大的计算资源。因此，必须选择合适的 3D 数据表示，并且必须将数据简化为易于处理的大小。问题是如何简化数据？随机点采样是一种常见的采样策略。但是，它对密度的变化很敏感。我们基于结合点云和体素数据的混合模型构建了一个采样模块。为了确定每个体素内的点之间的关系，该模块使用点的大小（点与对象中心之间的欧几里德距离）以及嵌入每个体素内的每个点之间的角度。通过利用从集合中的一个点开始并从已经选择的点中迭代地选择最远点的最远点采样（FPS），我们的方法具有在给定数量的质心内覆盖整个点集的优点，并且仍然保持关键点云和体素的优点是更好地表征包含在 3D 形状中的几何细节。进一步观察每个单元格中的点数不同，我们使用点量化方法来确保每个单元格具有相同的点数。这允许所有体素具有相同的特征大小向量，使 3D 卷积核更容易提取对象特征。我们展示了这些好处，并与 ModelNet10、ModelNet40 和 ShapeNetPart 数据集上的可靠基线进行了比较，