nips2017_pointnet++: Deep hierarchical Feature Learning on point sets in a Metric Space

Author

And pointnet is the same batch of authors, this is an improved version of Pointnet

Abstract

Pointnet does not capture the local structure introduced by metric spaces, it also limits its ability to recognize fine-grained categories and the ability to generalize complex scenes.

In this paper, a hierarchical neural network is applied recursively to gather the input points of nested partitioning.

By exploring the distance of metric space, the network structure designed in this paper can learn local features with the increase of the context scale.

The deeper observation found that the density of the point cloud is generally uneven, but the pointnet is trained on the point set of uniform sampling density, which results in great performance degradation, we propose that the new point-set learning method can adapt the characteristics of different scales.

Introduction

In this distance scale space, local areas may show different properties, such as different parts of its point cloud density is different

The basic idea of pointnet is to learn the spatial coding of each point and then aggregate the features of all points into global features, but this structure captures no local structure introduced by the measure. However, it is very important to explore the local structure in CNN, the traditional 2D CNN accepts the rules input, and constantly captures the characteristics of different scales in different layers, which can be better generalized to unseen cases by the constant abstraction of hierarchical structure.

In this paper, a hierarchical neural network is proposed to deal with the point cloud of hierarchical sampling.

Similar to the CNN structure, we first extract the local features from small areas, and then group to produce higher-level features, which are repeated until the entire point-gathering global feature is obtained.

To resolve two issues:

• How to divide this point cloud collection
• How to combine the local features of point gathering

These two problems are closely related, because the set of points is to let them have a common structure, so the local feature learner weight can be shared, similar to CNN, this article uses pointnet as a local feature learner

As a basic structural component, pointnet to extract features in a local point set, or to combine features into higher-level representations, so pointnet++ recursively uses pointnet to gather together at a nested input point

One question is how to divide the input:

Each local division is a domain sphere, with parameters that include the position and scale of the center point, which is sampled using the farthest point sampling (FPS) algorithm to more evenly cover the entire data set.

FPS: Select a random point first, then select the point farthest from the point as the starting point, and then continue the iteration, know how to select the number of required

Reference:

Https://www.cnblogs.com/li-yao7758258/p/8182846.html

It is essentially a layered version of pointnet. Each layer has three sub-stages: sampling, grouping, and pointneting. In the first stage, the centroid is selected, and in the second phase, the neighboring points around them (within a given radius) are created to create multiple sub-point clouds. They then give them a pointnet network and get a higher-dimensional representation of these sub-point clouds. They then repeat the process (sample centroid to find higher-order representations of their neighbors and pointnet to get a higher-dimensional representation). Use 3 of these network layers. A number of different aggregation methods at different levels were also tested to overcome the difference in sample density (this is a big problem for most sensors when the object approaches dense samples and is sparse in the distance). They have been improved on the prototype pointnet, and the accuracy rate on the MODELNET40 has reached 90.7%.

nips2017_pointnet++: Deep hierarchical Feature Learning on point sets in a Metric Space