1. 四个问题
-
解决什么问题
点云局部操作子的探索 -
用了什么方法解决
提出了GRA模块，基于此构建了RPNet完成点云分析。
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
-
效果如何
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 和68.2)
此外，在 parameters,computation saving， robustness，noises均有不错的表现。 -
还存在什么问题
？
2. 论文介绍
1. Introduction
对PointNet，PointNet++，RS-CNN，self-attention的分析问题，以及引出自己方法。（看别人分析问题，以及如何引出自己的方法，研究动机）
- 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)
( 有从邻域内预定义的 geometric relations（文中称为low-level relation）计算出中心点的特征，但是缺少semantic relations（high-level relations），semantic relations 从哪里来？ feature？)（However, RS-CNN is insufficient to learn semantic relations (high-level relations) for the lack of interaction between features）
- self-attention：self-attention能够补充这种 high-level relations，但是受限于 运算量和参数量
- 本文：The goal of this work is to extend grid-based self-attention to irregular points with a high-efficiency strategy。（提出bottleneck的GRA模块（既能学到 geometric shape又能学到semantic information），基于模块构建了宽的和深的网络：width (RPNet-W) and depth (RPNet-D)，宽的在分类上好，深的在分割上好，而且high-efficiency）
整个网络：
实验结果：
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. 参考资料
4. 收获
本文从分析PointNet，PointNet++,RS-CNN，self-attention出发，到引出自己的方法的分析过程值得学习。
- PointNet，PointNet++：没有从邻域交互
- RS-CNN：预定义的 low-level的 几何特征，没有high-level 的sementic infomation。
- self-attention：有sementic infomation
本文：GLR融合起来都有，learn from both low-level and high-level relations， concat起来，再用MLP去映射。还引入了cross-channel attention。
- Low-level Relation：预定义的几何特征
- high-level relations 从哪里来？ features？ features就代表语义信息？
