Shulou(Shulou.com)06/01 Report--

This article introduces how SO-Net is used for point cloud analysis. The content is very detailed. Interested friends can refer to it for reference. I hope it can help you.

Abstract

SO-Net, a permutation invariant network structure for deep learning of disordered point clouds, is proposed below. SO-Net simulates the spatial distribution of point clouds by constructing self-organizing maps (SOMs). Based on SOM, SO-Net performs hierarchical feature extraction for single point and SOM node, and finally uses a single feature vector to represent the input point cloud. The receptive field of the network can be adjusted systematically by performing a point-to-node KNN(k-nearest neighbor search). In tasks such as point cloud recognition reconstruction, classification, object segmentation and shape retrieval, our proposed network shows similar or better performance than state-of-the-art methods. In addition, due to the parallelism and simplicity of the proposed architecture, the training speed is much faster than that of existing point cloud recognition networks.

Introduction

After years of intensive research, convolutional neural networks (ConvNets) are now the basis for many of the most advanced computer vision algorithms, such as image recognition, object classification, and semantic segmentation. Despite ConvNets 'great success with 2D images, using deep learning on 3D data remains a challenging problem. Although 3D convolutional networks (3D ConvNets) can be applied to 3D data rasterized into voxel representations, most computations are redundant due to the sparsity of most 3D data. Furthermore, the performance of immature 3D ConvNets is largely limited by the loss of resolution and exponentially increasing computational costs. At the same time, the accelerated development of depth sensors and the huge demand for applications such as autonomous vehicles make efficient processing of 3D data a priority. The latest availability of 3D datasets including ModelNet [37], ShapeNet [8], 2D-3D-S [2] has increased the popularity of 3D data research.

To avoid the disadvantages of simple voxelization, one option is to explicitly exploit the sparsity of the voxel mesh [35, 21, 11]. While sparse designs allow for higher mesh resolution, their induced complexity and limitations make it difficult to implement large-scale or flexible deep networks.[30] Another option is to utilize scalable index structures, including kd-trees [4], octrees [25]. Deep networks based on these structures show encouraging results. Compared to tree-based structures, the point cloud representation is mathematically simpler and straightforward because each point is represented by only a 3-dimensional vector. Furthermore, with the help of the Structure of Motion (SfM) algorithm, point clouds can be easily acquired using popular sensors such as RGB-D cameras, LiDAR or regular cameras. Although point clouds are widely used and easily accessible, the task of identifying them remains challenging. Traditional deep learning methods such as ConvNets do not work because point clouds are spatially irregular and can be arranged arbitrarily. Because of these difficulties, few attempts have been made to apply deep learning techniques directly to point clouds, until recently PointNet [26].

Despite being a pioneer in applying deep learning to point clouds, PointNet still cannot adequately handle local feature extraction. PointNet++[28] was proposed to solve this problem by constructing a pyramidal feature aggregation scheme, but the point sampling and grouping strategy in [28] did not reveal the spatial distribution of the input point cloud. Kd-Net [18] constructs kd-trees from input point clouds and then performs hierarchical feature extraction from leaves to root nodes. Kd-Net explicitly exploits the spatial distribution of point clouds, but there are still limitations such as non-overlapping receptive fields.

In this paper, we propose SO-Net to solve problems in existing point-cloud-based networks. Specifically, SOM [19] was built to simulate the spatial distribution of the input point cloud, which makes it possible to perform hierarchical feature extraction on individual points and SOM nodes. Finally, the input point cloud can be compressed into a single feature vector. During feature aggregation, receptive field overlap is controlled by performing point-to-node k-nearest neighbor (KNN) search on SOM. Theoretically, SO-Net ensures that the order of input points remains unchanged through special network design and our permutation invariant SOM training. Applications of our SO-Net include point-cloud-based classification, autoencoder reconstruction, part segmentation, and shape retrieval, as shown in Figure 1.

The main contributions are as follows:

We design a permutation invariant network-SO-Net that explicitly exploits the spatial distribution of point clouds.

Hierarchical feature extraction can be performed by systematically adjusting receptive field overlap by performing a point-to-node KNN search on SOM.

We propose a point-cloud autoencoder as pre-training to improve network performance in various tasks.

Similar or better performance is achieved in a variety of applications compared to state-of-the-art methods, and training speeds are significantly faster.

About the point cloud analysis of the self-organizing network SO-Net is how to share here, I hope the above content can be of some help to everyone, you can learn more knowledge. If you think the article is good, you can share it so that more people can see it.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.