Deep residual neural network based PointNet for 3D object part segmentation
- PDF / 2,052,738 Bytes
- 15 Pages / 439.642 x 666.49 pts Page_size
- 11 Downloads / 211 Views
Deep residual neural network based PointNet for 3D object part segmentation Bin Li1
· Yonghan Zhang1 · Fuqiang Sun1
Received: 16 March 2020 / Revised: 11 July 2020 / Accepted: 12 August 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Point cloud segmentation is the premise and basis of many 3D perception tasks, such as intelligent driving, object detection and recognition, scene recognition and understanding. In this paper, we present an improved PointNet for 3D object part Segmentation, and named the proposed PointNet as Deep Residual Neural Network Based PointNet (DResNet-PointNet). The architecture of DResNet- PointNet was desigined based on the idea of residual networks. Residual networks can increase the depth of the DResNet-PointNet without network degradation. The depth of DResNet-PointNet is twice as deep as that of original PointNet model. Increasing the depth of DResNet-PointNet can improve its ability to express complex functions and generalization ability of complex classification problems, and achieve better approximation of complex functions, thus improving the accuracy of segmentation. The experimental results of part segmentation verify the feasibility and effectiveness of DResNet-PointNet. Keywords Point cloud · Point cloud segmentation · Deep residual neural network · PointNet
1 Introduction Due to the development of 3D sensors such as 3D laser scanning, we have more and more ways to acquire 3D image data. 3D image segmentation is the key step in tasks such as automatic drive [27], medical image processing [2, 4, 15, 26, 34], and virtual reality [20, 31]. Since the point cloud consists of a set of sparse and disordered 3D points, it is difficult to process the point cloud data directly by using deep neural networks such as Convolutional Neural Network (CNN) [16]. This is because convolution operations need to be performed on regular and sequential 2D or 3D models. At present, most of the existing point cloud processing methods are based on hand-crafted features. In [12], point cloud data were transformed into a two-dimensional spin image [11], and a histogram was computed from the spin image. Finally, objects were segmented from the scene by using clustering algorithm. Bin Li
[email protected] School of Computer Science, Northeast Electric Power University, Jilin, 132012, China
Multimedia Tools and Applications
In [1], minimal-z-value and contour-tracing algorithm [21] are used to segment the ground and buildings in the scene, respectively. Kalogerakis et al. [14] used the conditional Random Field (CRF) [17] to calculate the joint probability distribution of 3D point coordinates and labels of point cloud data, to achieve the purpose of image segmentation. Rusu et al. [29, 30] proposed Point Feature Histograms which describe the local geometry around each point in a point cloud dataset. In order to process point cloud data using deep neural networks, literatures [5, 25, 35, 37, 40] pre-process point cloud data into voxel models, while the literat
Data Loading...
