1. Four questions

1. What problem to solve
   Exploration of local operators of point cloud

2. What method has been used to solve
   Put forward GRA modular , Based on this, we build RPNet Complete point cloud analysis .

To this end, we propose group relation aggregator (GRA) to learn from both low-level and high-level relations. Compared with self-attention and SA, our designed bottleneck version of GRA is obviously efficient in terms of computation and the number of parameters. With bottleneck GRA, we construct the efficient point-based networks RPNet

3. What's the effect
   RPNet achieves state-of-the-art for classification and segmentation on challenging benchmarks（classification on ModelNet40 - 94.1, semantic segmentation on the datasets of ScanNet v2 and S3DIS (6-fold cross validation-70.8 and 68.2)
   Besides , stay parameters,computation saving, robustness,noises Good performance .

4. What's the problem
   ？

2. Paper introduction

1. Introduction

Yes PointNet,PointNet++,RS-CNN,self-attention Analysis of problems , And how to lead yourself .（ Watch others analyze problems , And how to lead out your own methods , Research motivation ）

• PointNet：For the ignorance of local structures
• PointNet++：However, this aggregator keeps learning on points independently, losing the sight of shape awareness
• RS-CNN：RS-CNN [33] computes a point feature from the aggregation of features weighted by predefined geometric relations (low-level relation) between the point Si and its neighbors N (Si) (shown in Fig. 1 middle)

( There are predefined... From the neighborhood geometric relations（ It's called low-level relation） Calculate the characteristics of the center point , But the lack of semantic relations（high-level relations）,semantic relations Where are from ？ feature？)（However, RS-CNN is insufficient to learn semantic relations (high-level relations) for the lack of interaction between features）

• self-attention：self-attention Can supplement this high-level relations, But limited by Operation quantity and parameter quantity

• this paper ：The goal of this work is to extend grid-based self-attention to irregular points with a high-efficiency strategy.（ Put forward bottleneck Of GRA modular （ Can learn geometric shape Can learn again semantic information）, Based on the module, a wide and deep network is constructed ：width (RPNet-W) and depth (RPNet-D), Wide ones are good in classification , Deep is good in segmentation , and high-efficiency）

The whole network ：

experimental result ：
classification on ModelNet40

semantic segmentation on the datasets of ScanNet v2 and S3DIS (6-fold cross validation).

4.4. Ablation Study

Inner-group relation function H

Aggregation function A.

Cross-channel attention.

4.5. Analysis of Robustness

Robustness to rigid transformation.

noises.

3. Reference material

4. Harvest

This paper analyzes PointNet,PointNet++,RS-CNN,self-attention set out , The analysis process leading to your own method is worth learning .

• PointNet,PointNet++： No interaction from neighborhood
• RS-CNN： Predefined low-level Of Geometric features , No, high-level Of sementic infomation.
• self-attention： Yes sementic infomation

this paper ：GLR It's all integrated ,learn from both low-level and high-level relations, concat get up , Reuse MLP Demapping . It also introduced cross-channel attention.

• Low-level Relation： Predefined geometric features
• high-level relations Where are from ？ features？ features It stands for semantic information ？