点云作为一种常见的三维数据表达和存储方式，被广泛运用于三维计算机视觉的应用中，如自动驾驶领域的点云分类、分割、重建等任务。近年来，深度神经网络由于具备强大的特征表达能力，常被用于处理和提取点云特征。然而，随着三维传感设备性能和存储能力的提升，高密度的点云数据带来了处理和存储方面的挑战，因此点云采样被广泛运用于各类任务中。但点云的传统采样算法往往只关注点云的几何结构，而不考虑采样后的相关任务，造成了任务的性能损失。如何在考虑任务的情况下设计任务驱动的自适应点云采样算法，成为了三维计算机视觉领域活跃的主题。针对这一问题，本文分别对于任务前的预采样、任务内部的层级采样这两个任务驱动的点云采样场景进行研究，获得了一定的成果。本文的主要研究内容和贡献如下：? 对于任务前的预采样场景，本文设计了一种面向下游任务的点云采样语义保持训练框架。现有的自适应点云采样框架使用端到端硬标签的方式训练点云采样网络，本文认为该框架没有充分利用完整点云的语义信息，造成了一定程度的语义丢失。本文在现有自适应点云采样框架的下采样分支基础上，增加了完整点云的监督分支，利用下游任务网络提取完整点云语义信息。为了充分利用完整点云语义信息，本文提出了两种级别的语义保持模块：标签级别的语义保持模块、特征级别的语义保持模块，分别使用软标签监督损失和特征对比损失来指导采样网络的训练。在多种点云任务上的实验验证了本文方法的有效性。? 对于任务内部的层级采样场景，本文设计了一种多尺度复用的点云自适应采样层。现有的一些自适应采样层专注于网络结构的设置，但只具备单尺度采样的能力，因此无法在多采样尺度上复用。而本文的采样模块可以处理任意密度点云的输入和输出，从而实现只引入单个模块的参数量，在点云任务网络的多级采样中共享参数；本文的采样模块的输入与输出只有点云的空间信息，不使用或改变任务网络中的特征提取模式，只替换其中的点云空间采样，具备更强的网络兼容性。在多种任务网络结构、多个下游任务上的实验证明了本文方法的有效性。

Point clouds, as a common representation and storage format for three-dimensionaldata, have been widely employed in applications of 3D computer vision, such as pointcloud classification, segmentation, and reconstruction tasks in the field of autonomousdriving. In recent years, deep neural networks, due to their powerful feature representation capabilities, have been frequently utilized for processing and extracting features frompoint clouds. However, with the improvement in the performance and storage capacityof 3D sensing devices, high-density point cloud data brings challenges in processing andstorage. Consequently, point cloud sampling has been widely utilized in various tasks.However, traditional point cloud sampling algorithms often only focus on the geometricstructure of the point cloud, without considering the tasks related to the sampled data,resulting in performance loss. Designing task-driven adaptive point cloud sampling algorithms considering the tasks has become an active topic in the field of 3D computervision. In addressing this issue, this paper investigates two task-driven point cloud sampling scenarios: pre-sampling before tasks and hierarchical sampling within tasks, andachieves certain results. The main research content and contributions of this paper are asfollows:? For the pre-sampling scenario, this paper proposes a point cloud sampling semanticpreserving training framework tailored for downstream tasks. Existing adaptivepoint cloud sampling frameworks train the point cloud sampling network usingan end-to-end hard label approach. This paper believes that this framework doesnot fully utilize the semantic information of the complete point cloud, leading toa certain degree of semantic loss. Building upon the downsampling branch of existing adaptive point cloud sampling frameworks, this paper introduces a supervision branch for the complete point cloud, utilizing downstream task networks toextract semantic information. To fully utilize the semantic information of the complete point cloud, this paper proposes two levels of semantic preservation modules:label-level semantic preservation module and feature-level semantic preservationmodule, using soft label supervision loss and feature contrast loss to guide the training of the sampling network. Experimental validation on various point cloud tasksconfirms the effectiveness of this approach.? For the hierarchical sampling scenario within tasks, this paper designs a multiscalereusable point cloud adaptive sampling layer. Existing adaptive sampling layersfocus on network structure but possess only single-scale sampling capability, thusunable to reuse across multiple sampling scales. The proposed sampling modulein this paper can handle input and output of point clouds with arbitrary densities,thereby sharing parameters across multiple sampling levels in point cloud task net-works with a single module. The input and output of the sampling module in thispaper only involve spatial information of the point cloud, without altering the fea-ture extraction pattern in task networks, only replacing the point cloud spatial sam-pling, hence possessing stronger network compatibility. Experimental results onvarious task network structures and multiple downstream tasks demonstrate the ef-fectiveness of this approach.