Network Security Internet Technology Development Database Servers Mobile Phone Android Software Apple Software Computer Software News IT Information

In addition to Weibo, there is also WeChat

Please pay attention

WeChat public account

Shulou

An example Analysis of the Innovation of PointNet

2025-01-18 Update From: SLTechnology News&Howtos shulou NAV: SLTechnology News&Howtos > Internet Technology >

Share

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

Today, I will talk to you about the example analysis of PointNet innovation, which may not be well understood by many people. in order to make you understand better, the editor has summarized the following content for you. I hope you can get something according to this article.

The following is not carried out in detail according to the order of the paper, but only a few innovative points of PointNet are briefly mentioned.

one

Aiming at the disorder of point cloud, Maxpooling is used as the symmetric function. The maximum pooling operation is to compare all the members, leaving the largest one behind and discarding the rest, so the maximum value will not change no matter how the order changes.

What does a symmetric function mean? For example, addition is a symmetric function, 1-2-3-4-10, in a different order, 2-4-3-1-10, no matter how the order changes, it will not affect the result.

two

For rigid body changes-align the network T-net. T-net can also improve performance, with two T-net plus regularization contributing 2.1%, but strangely enough, T-net is no longer seen in PointNet++ code (this point is not mentioned in the paper, there are questions on github, but the author did not reply).

three

MLP (multilayer perceptron, to put it simply, full connection layer) is used in the feature extraction stage. The operations used in this structure are only multiplication and addition, which are symmetric functions, so they will not be affected by sorting.

Experimental effect

1. Classification

The data set is ModelNet40, which contains the CAD of 40 kinds of objects, and the point cloud is obtained by sampling. The author does not list MVCNN here because it is not as accurate as it is, but later improved versions have exceeded it.

2. Object segmentation:

3. The comparison between the number of parameters and MVCNN: it is obviously dominant.

After reading the above, do you have any further understanding of the example analysis of PointNet innovation? If you want to know more knowledge or related content, please follow the industry information channel, thank you for your support.

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.

Views: 0

*The comments in the above article only represent the author's personal views and do not represent the views and positions of this website. If you have more insights, please feel free to contribute and share.

Share To

Internet Technology

Wechat

© 2024 shulou.com SLNews company. All rights reserved.

12
Report