Inventory clearing series , I always wanted to write , But I haven't read it very carefully , Here's the list first . There will be time later to explain each paper in detail .

Implicit expression

Implicitly express what you are doing ？
Make a list to see 3D Expression form ha ：

So the figure clearly explains the implicit expression , It's a 3D The shape is transformed into a decision surface to learn .
In short , Namely , In my submission 3D The surface of the , That is, the surface of the rabbit in the picture above is 0, Inside or outside is -1 or 1, So I just need to learn this hyperplane , That's it 3D Surface can . The problem is much simpler at once . Because in the end, we only need to predict that this point in space is 0、-1、 still 1.
Draw a picture as follows , It means implicit expression, learning is the black line y This part of the function , That is to say 3D The side of space . Red boxes and green circles are -1 and 1 The situation of , Not a surface .

Related papers

1. NIPS2016：Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling

Summary

• Put forward 3D-GAN, Generate 3D objects from probabilistic hidden space ; namely ：
  

• Explore the hidden space expression , In fact, it just explores the combination of hidden space features
  The overall architecture is as follows ：
  
  

• Of a single graph 3D restructure , Here's the picture

The detailed content

• The paper ：http://3dgan.csail.mit.edu/papers/3dgan_nips.pdf
• Project home page ：http://3dgan.csail.mit.edu/
• Code ：https://github.com/zck119/3dgan-release
• Translation reference ：https://blog.csdn.net/qq_39426225/article/details/101684526

2. ICML2018：Learning Representations and Generative Models for 3D Point Clouds

Summary

1. Yes 3 A generation model for in-depth study

•   Running on the original point cloud GANs
  

•   stay AEs Training in a fixed diving space GANs
  

•   Gaussian mixture model , The overall effect of this model is the best 
  

2. Proposed to match metrics：chamfer,coverage metric.

The detailed content

• The paper ：https://openreview.net/forum?id=r14RP5AUz
• Code :https://github.com/optas/latent_3d_points
• Reference reading ：https://blog.csdn.net/e2297192638/article/details/89299545

3. CVPR2019：Learning Implicit Fields for Generative Shape Modeling

Summary

Use implicit field As a decoder to do shape generation , Also by building surfaces to generate shapes . The benefits of this implicit expression , In fact, it is on the surface for 0, Inside or outside is -1 or 1, This is actually a problem of hyperplane fitting . The encoder can be universal CNN or PointNet.

The detailed content

• The paper ：https://arxiv.org/abs/1812.02822
• Project home page ：https://www.sfu.ca/~zhiqinc/imgan/Readme.html
• Code ：https://github.com/czq142857/IM-NET-pytorch

4. CVPR2020：Points2Surf: Learning Implicit Surfaces from Point Cloud Patches

Summary

The generation process of implicit surface can be limited , there SDF In essence, it is still saying the meaning of implicit expression ： It's on the surface 0, Inside or outside >0 or <0.

The detailed content

• The paper ：https://arxiv.org/abs/2007.10453、https://publik.tuwien.ac.at/files/publik_291233.pdf
• Code ：https://github.com/ErlerPhilipp/points2surf

5. CVPR2020：Neural Implicit Embedding for Point Cloud Analysis

Summary

This article is essentially through the construction of local neural networks （ELM） To enhance the expression of shape . Now that you have learned this shape , Then many subsequent tasks can be used . Classification and segmentation are mentioned in the article , So the topic is also called point cloud analysis .

The detailed content

• The paper :
  https://openaccess.thecvf.com/content_CVPR_2020/papers/Fujiwara_Neural_Implicit_Embedding_for_Point_Cloud_Analysis_CVPR_2020_paper.pdf

• Code :
  https://github.com/kentfuji/NeuralEmbedding