利用早停可微架构搜索的三维点云模型分类

针对现有三维点云分类网络采用人工设计费时费力的问题，提出早停可微架构搜索(early-stopping differentiable architecture search,ES-DARTS)算法。利用从人工设计网络架构中提取到的先验知识，预定义一个包含高效候选操作的搜索空间，可快速搜索出适用于三维模型分类任务的高性能网络模型;通过追踪网络搜索阶段各候选操作的权重变化，找出跳跃连接操作在双重优化过程中发挥不公平竞争作用的临界点并在此处停止搜索，以保证各候选操作之间的稳定性，解决DARTS算法搜索过程中易出现性能崩溃的问题。提出的算法在ModelNet40数据集上达到了93.2%的识别准确率，比当前人工设计的主流网络具有更高的识别准确率。

3D point cloud model classification based on early-stopping differentiable architecture search

The problem with most of nowadays 3D point cloud model recognition and classification networks is that they are manually designed and their design processes are time-consuming, laborious and sometimes error-prone. To solve this problem, we propose an early-stopping differentiable architecture search (ES-DARTS) algorithm. Firstly, extracting the prior knowledge from the manually designed network architecture, we predefine a search space containing efficient candidate operations, which can quickly search high performance network architecture suitable for 3D model recognition and classification tasks. Secondly, an early stop strategy is proposed to prevent the performance of the search process in DARTS algorithm from collapsing. We find out, by tracking the weight change of each candidate operation during the network search stage, that the jump connection operation plays an unfair competitive role in the double optimization process. Stopping the search here ensures the stability between the candidate operations and allows us to search for a cell structure with superior performance. The accuracy rate of 93.2% is achieved on the ModelNet40 dataset. Compared with the current mainstream artificially designed network, the recognition accuracy of the network in this paper is not only higher, but also has stronger robustness.