Improvement of pointnet++

1、PATs: Point Attention Transformers
The paper ：Modeling Point Clouds with Self-Attention and Gumbel Subset Sampling

2019CVPR, Shanghai Jiaotong University MoE Laboratory and Huawei Noah's Ark Laboratory

This paper mainly improves PointNet++ Medium FPS Part of , Make the selected point more able to deal with outer points , take down sampling The selected point is attention score On the big point . A vivid comparison is shown below ：

Next, I will introduce the two modules proposed in this paper .

GSA：Group Shuffle Attention
This part is mainly about self attention mechanism .

This article uses Scaled Dot-Product attention（ The blog above also mentioned ）, Concrete GSA See the formula in the paper for details （7）（8） and （11）.

For point x, The corresponding eigenvector is y, adopt y=GSA(f) Transformation of , A new eigenvector of points is obtained y. stay GSA In the process , Reference resources (7) and (8), Still in use f Internal information , And MLP In fact, the effect is the same , Is to calculate new features .

GSS：Gumbel Subset Sampling
Mainly is to Pointnet++ Medium FPS Use instead Gumbel Softmax Calculate the importance of each point , Then choose according to this probability down sampling The point of .

Network structure

From the perspective of network structure ,GSA The Internet is similar to Pointnet++ Medium MLP,GSS amount to FPS.

Moments
The paper ：Momenet: Flavor the Moments in Learning to Classify Shapes

The method proposed in this article is simple , Basically, it is to increase the input of the network . Input the network by the coordinates of the points (x,y,z) Add to (x,y,z,x2,y2,z2,xy, xz, yz). The basic network is as follows ：

2nd Order-Layer
You can see from the details 2nd Order-Layer In fact, with Pointnet++ Very similar , Nothing more than artificial construction 2nd Order Characteristics of .

Concrete Spatial-Trans We also added 2nd Order-Layer.

The final effect proves , stay ModelNet40 On the dataset , The effect is very good .

reference ：https://blog.csdn.net/wqwqqwqw1231/article/details/104183984