基于邻域信息编码的点云语义分割

Point Cloud Segmentation Based on Neighboring Information Encoding

摘要: 由于点云这种数据在各个领域的广泛应用，如何对其进行充分的利用就成了亟需解决的问题。良好的点云的语义分割结果可以为点云下游应用任务提供良好的基础。对于语义分割任务而言，全局信息和局部信息都是必要的，都发挥着重要的作用。基于局部特征越好最终的分割效果也会越好这样一个基本事实设计了一个名为邻域信息编码的模块，这个模块在进行局部特征提取的时候将中心点与中心点周围的邻域点之间的方向信息充分利用。原始的输入特征被认为是空间特征和附加特征(例如：颜色和深度)的组合，附加特征通过共享多层感知机进行处理，空间特征将会通过所提出的邻域信息编码模块来提取每一个中心点的邻域特征，处理之后的两部分特征将会被拼接起来作为最终的局部输出特征。从在公开数据集SS3DIS上的实验结果来看，所提出的模型在分割任务上有良好的表现，它的分割结果比当前最好的基线模型高了1.8%。

Abstract: How to take full advantages of point cloud is an urgent task due to the universal application of point cloud data format. Superb semantic segmentation of point cloud can be the solid basis of subsequent task of point cloud application. For semantic segmentation task, both the global and local feature play significant role. Based on the basic truth that the better the local feature, the better the segmentation performance, the neighboring feature encoding module is elaborated, this module aims to make the most of direction feature between center point and its neighboring points during the local feature extraction process. Original input feature is considered as concatenation of spatial information and plus feature (e.g.: RGB and depth), the plus feature will be handled by shared MultiLayer Perceptron, the spatial information will be sent to proposed neighboring feature encoding module to capture neighboring feature of every center point, output of two part will be concatenated as final local feature. From the perspective of result of experiment on public dataset S3DIS, the proposed model has excellent performance on feature extraction, its performance excels SOTA baseline model with 1.8% mIoU.