Finally, I can simply write the guide to this piece , It's been delayed for more than a week . I hope I can help you ;）

1. Background knowledge

Point cloud retrieval （point cloud retrieval） vs Scene recognition （place recognition）

Point cloud retrieval , It's actually very intuitive , It's like Image retrieval is the same , Map search , It's just a little cloud search . With the data from 1D–>2D–>3D The development of , Such a retrieval system is naturally needed .

Scene recognition You can view it as Point cloud retrieval An example of , Or an application scenario .
Scene recognition is mainly aimed at automatic driving , It can identify whether the current scene is in the planned route and judge .

2. Definition

Suppose the database set is M, The point cloud queue is Q, Our goal is to retrieve the point cloud queue Q Each of these point clouds is in the data set M The one with the most similar structure . The design task is a little ：

• How to express a pair of point clouds ？
• How to measure similarity ？

3. traditional method

Step 3: datum line ：

1. feature extraction
2. Feature code
3. matching
   

Refer to the process for details ：
An evaluation of feature encoding techniques for non-rigid and rigid 3D point cloud retrieval

Other relevant references ：

1. A survey of content based 3d shape retrieval methods 2004
2. local descriptors: A comprehensive performance evaluation of 3d local feature descriptors（ Traditional local feature descriptors ）
3. mesh retrieval：Feature encoding of spectral signatures for 3d non-rigid shape retrieval
4. The bag of words approach for retrieval and categorization of 3d objects
5. 3D point cloud retrieval with bidirectional feature match IEEE Access 2019 （ matching ）
6. 3D point cloud retrieval with bidirectional feature match （ matching ）

4. Methods based on deep learning

Existing history ：

Introduction to individual methods ：

• PointNetVLAD： Pioneer work , Mainly combined with PointNet+NetVLAD+ Quad user function
• PCAN： Use PointNet++ The idea of grouping to improve feature expression ;
• LPD-Net： Features use learning based methods + Hand designed , At the same time, the graph structure is used to improve the effect , High cost
• DH3D： Use 6 freedom , Focus more on SLAM Direction , So there is no specific reading
• LPD-AE： yes LPD-Net Subsequent versions , Use AE, among Encoder For feature extraction ,Decoder For refactoring , Two tasks are realized at the same time ： The extracted features are used to retrieve ; Refactoring is used to reproduce features to point clouds ;
• DAGC： Mainly used DGCNN Graph convolution of , And two way attention.
• EPC-Net： Not yet .Efficient 3D Point Cloud Feature Learning for Large-Scale Place Recognition
• SOE-Net： Yes PointShift The thought of that article has directional characteristics .
• MinkLoc3D： Minkovs base point cloud convolution is used to simplify feature expression , At the same time Geo Mean Pool method replaces NetVLAD, Is the most effective method at present
• PPT-Net： Pyramid architecture + The graph structure , The best effect
• TransLoc3D： And PPT-Net Similar thinking , Yes Transformer Thought ,arXiv The latest article .

The follow-up that has not been explained in detail is slowly increasing .

5. Other reference ：

1. Related meetings ： 3D oject retrieva （3DOR）
2. Track：14 years of Retrieval Contest (SHREC) track
3. Relevant latest list ：https://github.com/kxhit/awesome-point-cloud-place-recognition