摘要
目前人们对二维图像的研究已经取得了非常好的结果,然而随着深度学习的发展,研究正在逐步由二维向三维数据发展,并且应用领域越来越广泛,比如自动驾驶领域的三维场景建模,VR虚拟环境等。对三维数据的研究也逐渐实现了由有序输入到无序输入的过度并且取得了很高的成绩。Point Net则是第一个突破点云数据无序性输入的深度神经网络,值得人们深入的研究和借鉴。但是目前对破损和遮挡的点云数据问题还有待研究。本文着重对Point Net进行了深入研究并对点云数据进行了攻击和测试。测试发现当点云数据dropout约75%以后物体识别准确率显著下降,overlap12. 5%以后准确率也下降了近4个点,值得后续深入的研究和攻克。
At present,people’s research on two-dimensional images has achieved very good results.However,with the development of deep learning,research is gradually developing from two-dimensional to three-dimensional data,and the application fields are more and more extensive,such as three-dimensional scenes modeling,automatic driving,VR virtual environment,etc.The research on 3D data has gradually realized the excessiveity from ordered input to disordered input and achieved high results.PointNet is the first deep neural network to break through the unordered input of cloud data,which is worthy of in-depth research and reference.However,the current problem of point cloud data for damage and occlusion remains to be studied.This article focuses on PointNet’s in-depth research and attack and test point cloud data.The test found that when the point cloud data dropout is about 75%,the object recognition accuracy rate drops significantly.After the overlap12.5%,the accuracy rate also drops by nearly 4 pointswhich is worthy of further research and conquer.
作者
王胜文
张彬
孙菁聪
WANG Sheng-wen;ZHANG Bin(Science School,Communication University of China,Beijing 100024,China)
出处
《中国传媒大学学报（自然科学版）》
2019年第3期51-57,共7页
Journal of Communication University of China：Science and Technology
关键词
深度神经网络
三维点云
点云分类
语义分割
数据破损
deep neural network
3D point cloud
object classification
semantic segmentation
data breakage
