1. Four questions

1. What problem to solve

self-supervised learning task for deep learning on raw point cloud data

2. What method has been used to solve

a neural network is trained to reconstruct point clouds whose parts have been randomly rearranged.While solving this task, representations that capture semantic properties of the point cloud are learned
agnostic of network architecture（ It has nothing to do with the network architecture ）

3. What's the effect

（ I didn't understand this paragraph ）A linear SVM is trained on the representations learned in an unsupervised manner on the ShapeNet dataset. ？

4. What are the problems

such Change the position at will and predict the original position Of Agent task , Really ？（ however There seems to be less lattice ,3x3）

2. Paper introduction

While solving this task, representations that capture semantic properties of the point cloud are learned

architecture-agnostic（ Not related to network architecture ）

In this paper we propose a self-supervised method that learns powerful representations from raw point cloud data.
Our method works by training a neural network to reassemble point clouds whose parts have been randomly displaced.
The key assumption of the proposed method is that learning to reassemble displaced point cloud segments is only possible by learning holistic representations that capture the high-level semantics of the objects in the point cloud.（ The key assumption of this method is , Only by learning to capture the overall representation of the high-level semantics of objects in the point cloud , To learn to reassemble the replaced point cloud segment .）

specific working means ：

5 Discussion

In all the experiments , The representation learned by our proposed method is proved to be effective . It makes us believe that , For point clouds , image [7,21] As discussed in the image domain , The simple solution to rebuilding the input task is not a meaningful problem （ That is to say , The agent task cannot be too simple ？ The task is more complicated , Can learn better results ？）

3. Reference material

4. Harvest

Auto-encoder And GAN What's the difference? ？

A new agent task ： Upset The order （3x3）, Predict the original location

5 Discussion

In all the experiments , The representation learned by our proposed method is proved to be effective . It makes us believe that , For point clouds , image [7,21] As discussed in the image domain , The simple solution to rebuilding the input task is not a meaningful problem （ That is to say , The agent task cannot be too simple ？ The task is more complicated , Can learn better results ？）