深圳大学学报理工版

为解决PointNet++精度较低、耗时较长,且对输入点的噪声敏感的缺陷,引入一种高效的k维树(k-dimensional tree, kd tree)邻域查询方法,通过构建kd tree查找查询点周围指定半径内的近邻点,构造局部区域集,完成在PointNet++分组层上的局部特征提取.针对原网络训练过程中存在的过拟合问题,引入随机失活(dropout)正则化,减少网络收敛训练的时间.在Ubuntu14.04系统下搭建TensorFlow的图形处理器深度学习环境,并在ModelNet40数据集上进行训练和测试.实验结果表明,分别为当查询半径为0.1、0.2和0.3时,该查询方法的分类准确率分别为91.1%、92.1%和94.3%,皆优于PointNet++方法,且网络训练用时更短.改进后的结构在斯坦福三维语义分析数据集(Stanford 3D semantic parsing dataset)上进行语义分割实验平均交并比(mean intersection over union, MIoU)达57.2%,且其对于遮挡物体的鲁棒性更高.

PointNet++ is one of the effective deep learning methods to deal with the point cloud classification. However, there are some problems in this method including low precision, being time-consuming and sensitive to the noise of input points. In order to deal with these problems, we propose an efficient neighbor query method in which a k-dimensional tree(kd tree)structure is constructed to find the neighbor points within a specified radius around the query points, and the local features are extracted at the grouping layer of PointNet++. Aiming at the overfitting problem existing in the original network training process, we introduce the dropout regularization and thus reduce the training time of network convergence. The experiment environment is TensorFlow framework under Ubuntu14.04 system. The training and testing experiments are conducted based on ModelNet40 dataset. The classification accuracy of this method reaches 91.1%, 92.1% and 94.3% when the query radius is 0.1, 0.2 and 0.3, which is higher than PointNet++, and the kd tree method takes less time than the original method. At the same time, the improved structure has better semantic segmentation performance on the Stanford 3D semantic analysis dataset(mean intersection over union reaches 57.2%), which demonstrates the proposed method being more robust to occlusion objects.

引言
1 邻域查询算法
2 dropout优化
3 实验及分析
4 结语

图1 集合提取层的结构<br/>Fig.1 Structure of set abstraction level

图1 集合提取层的结构
Fig.1 Structure of set abstraction level

图2 三维kd tree结构
Fig.2 3D kd tree

图3 本研究改进算法的训练日志-准确率变化<br/>Fig.3 (Color online)Training log-accuracy variation of the kd-tree method

图3 本研究改进算法的训练日志-准确率变化
Fig.3 (Color online)Training log-accuracy variation of the kd-tree method

表1 不同邻域查询方法的分类准确率<br/>Table 1 Effects of different neighborhood options

表1 不同邻域查询方法的分类准确率
Table 1 Effects of different neighborhood options

图4 场景语义分割定性对比结果<br/>Fig.4 (Color online)Qualitative comparison of scene semantic segmentation results

图4 场景语义分割定性对比结果
Fig.4 (Color online)Qualitative comparison of scene semantic segmentation results

[1] 陈艳雷.基于逆向工程的扫描点云数据预处理技术研究[D].郑州:河南工业大学,2018.
[2] BU Shuhui, LIU Zhenbao, HAN Junwei, et al. Learning high-level feature by deep belief networks for 3-D model retrieval and recognition[J]. IEEE Transactions on Multimedia, 2014, 16(8): 2154-2167.
[3]XIE Jin, DAI Guoxian, ZHU Fang, et al. Deepshape: deep-learned shape descriptor for 3D shape retrieval[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(7): 1335-1345.
[4] SU Hang, MAJI S, KALOGERAKIS E, et al. Multi-view convolutional neural networks for 3D shape recognition[C]// IEEE International Conference on Computer Vision. Santiago, Chile: IEEE, 2015: 945-953.
[5] SHI Baoguang, BAI Song, ZHOU Zhichao, et al. Deep-Pano: deep panoramic representation for 3-D shape recognition[J]. IEEE Signal Processing Letters, 2015, 22(12): 2339-2343.
[6] SINHA A, BAI Jing, RAMANI K. Deep learning 3D shape surfaces using geometry images[C]// The 14th European Conference on Computer Vision. Amsterdam: Springer, 2016: 223-240.
[7]WU Zhirong, SONG Shuran, KHOSLA A, et al. 3D ShapeNets: a deep representation for volumetric shapes[C]// IEEE Conference on Computer Vision and Pattern Recognition. Boston, USA: IEEE, 2015: 1912-1920.
[8] XU Xu, TODOROVIC S. Beam search for learning a deep convolutional neural network of 3D shapes[C]// The 23rd International Conference on Pattern Recognition. Cancun, Mexico: IEEE, 2016:3506-3511.
[9]QI C R,SU Hao, KAICHUN M, et al. PointNet: Deep learning on point sets for 3D classification and segmentation[C]// IEEE Conference on Computer Vision and Pattern Recognition. Hawaii, USA: IEEE, 2017: 77- 85.
[10]QI C R, YI Li, SU Hao, et al. PointNet++: deep hierarchical feature learning on point sets in a metric space[C]// Advances in Neural Information Processing Systems 30. California, USA:[s.n.], 2017: 5105-5114.
[11] BENTLEY J L, FRIEDMAN J H. Data structures for range searching[J]. ACM Computing Surveys, 1979, 11(4):397- 409.
[12] CHENG Dongcai, MENG Gaofeng, XIANG Shiming, et al. FusionNet: edge aware deep convolutional networks for semantic segmentation of remote sensing harbor images[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2017, 10(12): 5769-5783.
[13] FRANCIS E, THEODORA K, ALEXANDER H, et al. Exploring patial context for 3D semantic segmentation of point clouds[C]// IEEE International Conference on Computer Vision Workshop. Venice, Italy: IEEE, 2017: 716-724.
[14] BOULCH A, SAUX B L, AUDEBERT N. Unstructured point cloud semantic labeling using deep segmentation networks[C]// Eurographics Workshop on 3D Object Retrieval. Lyon, France: The Eurographics Association, 2017: 17-24.

备注

引言

1 邻域查询算法

2 dropout优化

3 实验及分析

4 结语

期刊信息

备注

引言

1 邻域查询算法

2 dropout优化

3 实验及分析

4 结 语

期刊信息

4 结语