面向点云识别网络的显著图生成  

Saliency maps for point cloud recognition models

作　　者：梁奥 张浩[1,2,3,4] 花海洋[1,2] LIANG Ao;ZHANG Hao;HUA Haiyang(Key Laboratory of Opto-Electronic Information Processing,Chinese Academy of Sciences,Shenyang 110016,China;Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China;Institutes for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang 110169,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院、光电信息处理重点实验室,辽宁沈阳110016 [2]中国科学院沈阳自动化研究所,辽宁沈阳110016 [3]中国科学院、机器人与智能制造创新研究院,辽宁沈阳110169 [4]中国科学院大学,北京100049

出　　处：《光学精密工程》2023年第8期1188-1201,共14页Optics and Precision Engineering

基　　金：中科院创新基金资助(No.E01Z040101)。

摘　　要：点云结构上的特殊性质,导致解释其深度模型学习特征的结果存在困难。提出了一种获得点云目标识别模型显著图的方法,首先在点云空间中随机释放若干自由因子并输入到模型中,然后根据设计的贡献度评价指标,基于梯度下降使骨干网络输出的池化特征尽可能偏离目标点云识别过程中输出的池化特征并更新因子位置。迭代后的因子无法参与识别过程,其对模型的预测“零贡献”,将目标点云中的点移动到这些因子的位置后对识别结果的影响与丢弃该点完全相同。点的移动过程可微,最后可根据梯度信息获得显著图。本文的方法在ModelNet40数据集上生成PointNet模型的显著图,相较于用点云中心生成显著图的方法,理论依据更强且适用的数据集更多。移动点至“零贡献”因子位置后对模型的影响较移动点至点云中心与丢弃点更相似。按本文的方法丢弃点使模型精度下降得更快,在仅丢弃100个点的情况下,模型的OA(overall accuracy)由90.4%下降至81.1%。同时经DGCNN和PointMLP评估,该显著性结果具有良好的通用性。该方法生成的显著性分数精度更高,且由模型驱动不含任何假设,适用于绝大多数点云识别模型和数据集,其显著性分析结果对目标识别网络的搭建与数据增强具有指导意义。The specific properties of point cloud structures lead to difficulties in the interpretation of fea⁃tures learned from deep neural networks(DNNs).In this study,a method is proposed to obtain saliency maps for point cloud target recognition models.First,a number of free factors are randomly released in the point cloud space and input into the model.Then,based on the designed contribution evaluation in⁃dex,the pooled features output by the backbone are deflected as much as possible using the target point cloud recognition process based on gradient descent,and the factor positions are updated.The iterated fac⁃tors do not participate in the recognition process,thus contributing"zero"to the prediction of the model such that moving the points in the target point cloud to the positions of these factors has exactly the same effect on the recognition result as dropping the points.The process of moving the points is differentiable,and the saliency maps can be obtained from the gradient information.The saliency maps for PointNet were generated on ModelNet40 using the proposed method.This method has a strong theoretical basis for gen⁃erating saliency maps and is applicable to more number of datasets compared with the method using point cloud centers.The effect of shifting points to the no-contribution factor position is more similar to dropping points than shifting points to the center of the point cloud.In this study,dropping points by the saliency scores rapidly reduced the overall accuracy of the model from 90.4%to 81.1%with only 100 points dropped.Meanwhile,the saliency scores displayed good generality as evaluated using deep graph convolu⁃tional neural network(DGCNN)and PointMLP.The proposed method is driven by the model without prior assumptions and generates significance scores with higher precision;it is applicable to most point cloud recognition models.The results of saliency analysis are instructive for the construction of target rec⁃ognition networks and data augmentation.

关 键 词：点云 显著图 目标识别 深度学习 可解释性 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]