3D点云深度学习-浅谈点云分割

先说一点题外话

读研究生三年,我开始研究图像检测一年,但我还不懂,然后实验室都开始做点云,然后我切换到点云方向两年。我没有取得太大的成就,因为我觉得我才刚开始学习了两年。我的大部分研究都是现成的代码。即使输出了一些结果,我也没有真正改变大网络,也没有写出一整套代码。但从长远来看,我对这个方向有点熟悉。现在我已经毕业了,我没有机会再做这个方向了。我会根据我自己对联系路线的了解,把它介绍给一些刚刚接触到我的同学。我所说的不一定是正确的,但我会整理出来,等我有空的时候再加进去。

[En]

Three years as a graduate student, I began to study image detection for a year, but I didn’t understand it yet, and then the lab all started to do point cloud, and then I switched to the point cloud direction for two years. I didn’t achieve much, because I felt that I had just started after two years of study. And most of my studies are ready-made codes. Even if some of the results are output, I haven’t really changed the big network or written a whole set of code. But in the long run, I am a little familiar with this direction. Now that I have graduated, I have no chance to do this direction again. I will introduce it to some students who have just come into contact with me according to my own understanding of the contact route. What I have said is not necessarily correct, but I will sort it out and add it when I am free.

背景

1.点云是什么?
官方解释:点云是大量点的集合,这些点表示目标在同一空间参照系中的空间分布和表面特征。

[En]

Official explanation: a point cloud is a collection of massive points that express the spatial distribution and surface characteristics of the target in the same spatial reference frame.

实际上:点云是一组三维点(x,y,z)的集合。
2.深度学习是什么?
我不知道,只是各种各样的文件,就像去上班一样,都结束了。

[En]

I don’t know, it’s just all kinds of papers, just like going to work, it’s over.

3.点云深度学习的难点在哪里?
事实上,图像中的深度学习已经取得了很好的效果,商业模型随处可见,而点云最大的困难是如何解决点云的无序问题,即同一对象的这么多点按不同的顺序排列。结果应该是一样的,目前在思考如何更好地解决这个问题,当然,点云中存在遮挡、噪声点等因素。

[En]

In fact, deep learning in the image has achieved very good results, commercial models can be seen everywhere, and the biggest difficulty of point cloud is how to solve the problem of disorder of point cloud, that is, so many points of the same object are arranged in different order. the results should be the same, currently thinking about how to better solve this problem, of course, there are occlusion, noisy points and other factors in the point cloud.

4.为什么点云深度学习最近比较火?
一方面是深度学习目前的学习能力比较强,另一方面,3D传感器发展的很多,更加接近真实世界,而且近期出现了很多3D应用,可能再有就是电云可能相对容易发论文一些吧,哈哈哈~
5.点云深度学习领域包括哪些领域的学习?
图像卷积神经网络、数据处理、可视化、TF/Torch框架
6.点云的研究方向有哪些?
点云检测、点云分割、点云识别、点云补全、点云数据增强、点云数据集制作及Benchmark

点云数据集

深度学习任务必须没有数据集,模型需要大量的训练样本,当然也可以使用深度相机或激光雷达进行实时采集和处理,也就是模型训练后的应用层,因为你参与了一篇数据集评论文章的准备,所以数据集比较多,点云领域有各种各样的数据集,因为每个人都有不同的采集设备和不同的采集策略。一些后期的处理和标记格式是不同的,顺便说一句,他们采集的场景的比例和类型也不同,所以我们可以研究某种场景类型的点云方向。例如,专门从事点云对象识别、室内场景点云分割、室外场景点云分割等,介绍了一些常用的数据集。

[En]

Deep learning tasks must be without data sets, the model requires a large number of training samples, of course, you can also use a depth camera or lidar real-time acquisition and processing, that is the application level after the model training, because you have participated in the preparation of a data set review article, so there are more data sets, and there are a variety of data sets in the point cloud field, because everyone has different acquisition equipment and different acquisition strategies. Some later processing and marking formats are different, by the way, the scale and type of the scene they collect are also different, so we can study the point cloud direction for a certain scene type. for example, specializing in point cloud object recognition, indoor scene point cloud segmentation, outdoor scene point cloud segmentation and so on, introduce some commonly used data sets.

1.Semantic3D
这应该是最经典的室外场景电云分割数据集,一台固定的激光雷达一直在扫描周围场景获取数据集,精度还是不错的,很多论文都是基于这个数据集,他提供官网提交你跑步的结果。

[En]

This should be the most classic outdoor scene electric cloud segmentation data set, a fixed lidar has been scanning the surrounding scene to get the data set, the accuracy is still good, many papers are based on this data set, and he provides the official website to submit the results of your run.

3D点云深度学习-浅谈点云分割
工程地址:http://www.semantic3d.net/
3D点云深度学习-浅谈点云分割
2.S3DIS
S3DIS数据集是斯坦福大学开发的带有像素级语义标注的语义数据集。室内场景点云数据集,一般是3D相机或者iPad什么的进行采集。
工程地址:http://buildingparser.stanford.edu/dataset.html
3D点云深度学习-浅谈点云分割
3.SemanticKITTI
连续多帧点云数据集,好像就是kitti数据集后面又经过处理得到的,汽车采集。
工程地址:http://www.semantic-kitti.org/index.html
3D点云深度学习-浅谈点云分割
3D点云深度学习-浅谈点云分割

4.SensatUrban
户外景点云数据集,真的是大体上、无人机分区域采集的数据集。

[En]

Outdoor scenic spots cloud data set, is really large, UAV sub-regional collection of data sets.

项目地址:http://point-cloud-analysis.cs.ox.ac.uk/
数据集挑战赛网址:https://competitions.codalab.org/competitions/31519#results

3D点云深度学习-浅谈点云分割
5.Modelnet
竟然忘记了最经典的分类数据集-ModelNet,该系列延伸出很多数据集,包括ModelNet-10和ModelNet-40,都是最开始接触点云领域用到的数据集。
网站:http://modelnet.cs.princeton.edu/#
6.ShapeNet
这是一个经典的、早期的数据集。
[En]

It’s kind of a classic, early data set.

; 点云深度学习模型

深度学习模型目前主要分为几个方向:基于原始点云、基于图像、基于图神经网络、基于3DCNN的方向去做,主要的区别就是将点云数据转化为不同的数据源去处理。介绍几个比较经典的网络模型。
1.PointNet/PointNet++
最经典的点云处理网络给人的感觉是,它现在被别人当成了一个模块。我不得不说,它非常容易使用。

[En]

The most classic point cloud processing network feels that it is now used as a module by others. I have to say that it is very easy to use.

3D点云深度学习-浅谈点云分割
2.RandLA-Net
牛津大学胡庆拥CVPR2020的一篇文章,也是我毕设跑的最多的代码,SensatUrban也是这个作者提出数据集,并且也举办了挑战赛
3D点云深度学习-浅谈点云分割
3.PointCNN
4.PointConv
5.

相关资料

点云PCL

结语

周末在公司免费整理一下,有时间补充一下,有很多内容,先列出一个小框架~

[En]

At the weekend in the company free to sort out a little bit, have time to add, there are a lot of content, first list a small framework ~

Original: https://blog.csdn.net/qq_41986166/article/details/119486007
Author: 七小琦
Title: 3D点云深度学习-浅谈点云分割

原创文章受到原创版权保护。转载请注明出处:https://www.johngo689.com/496787/

转载文章受原作者版权保护。转载请注明原作者出处!

(0)

大家都在看

亲爱的 Coder【最近整理,可免费获取】👉 最新必读书单  | 👏 面试题下载  | 🌎 免费的AI知识星球