当前位置:网站首页>【点云压缩】Sparse Tensor-based Point Cloud Attribute Compression
【点云压缩】Sparse Tensor-based Point Cloud Attribute Compression
2022-06-12 02:44:00 【Jonathan_Paul 10】
Sparse Tensor-based Point Cloud Attribute Compression
该文运用了稀疏卷积,将多个稀疏卷积堆积成VAE的基本框架;用这个基本框架来做learning-based的方法。值得一提的是,本文是用在属性编码上的,而learning-based用在属性编码上很少。
Ideas
Background
Related work
End-to-End (E2E) Learning
对一个E2E的架构,一般会分为两个编码器:一个为主要的编码器(Main codec)和超先验的编码器(Hyper codec)。
在超先验编码器中,它会被Hyper Encoder提取先验信息,之后用Hyper Decoder来解码得到main codec的均值与方差 ( μ i , σ i ) (\mu_i,\sigma_i) (μi,σi)。
Sparse convolution
要解决的是对稀疏数据进行更有效的特征提取问题。我们知道,点云是具有稀疏性的,如果运用传统的卷积计算,花销势必会很大。因此,人们使用稀疏卷积来有效地计算稀疏数据,而并非对图片像素点(2D)和空间的体素点一一扫描。[1]
Quantization and Rate-Distortion
量化的粗糙程度决定了最终模型表现的好坏(即失真 D D D的多少)。而码率 R R R直接决定了量化的粗糙程度。
Model
Definition
首先需要定义一个Sparse Tensor: { C → , F → } \{\overrightarrow{\mathbf{C}}, \overrightarrow{\mathbf{F}}\} { C,F}。其中, C → = { ( x i , y i , z i ) ∣ i ∈ [ 0 , N − 1 ] } \overrightarrow{\mathbf{C}}=\left\{\left(x_{i}, y_{i}, z_{i}\right) \mid i \in[0, N-1]\right\} C={ (xi,yi,zi)∣i∈[0,N−1]}、 F → = { ( R i , G i , B i ) ∣ i ∈ [ 0 , N − 1 ] } \overrightarrow{\mathbf{F}}=\left\{\left(R_{i}, G_{i}, B_{i}\right) \mid i \in[0, N-1]\right\} F={ (Ri,Gi,Bi)∣i∈[0,N−1]}

在该模型框架中,熵编码模型将会被使用。而针对熵编码模型,作者沿用了[2]中的先验知识,以提升熵编码模型的表现。
Reference
[1] https://towardsdatascience.com/how-does-sparse-convolution-work-3257a0a8fd1
[2] VARIATIONAL IMAGE COMPRESSION WITH A SCALE HYPERPRIOR
边栏推荐
- 利用ssh公钥传输文件
- Maya foreground rendering plug-in Mel scripting tool
- How to make div 100% page (not screen) height- How to make a div 100% of page (not screen) height?
- In 2022, don't you know the difference between arrow function and ordinary function?
- cupp字典生成工具(同类工具还有crunch)
- SwiftyJSON解析本地JSON文件
- Summary of force deduction solution 436- finding the right interval
- Abaqus中批量对节点施加集中力荷载
- Force deduction solution summary 868- binary spacing
- 力扣解法汇总358-迷你语法分析器
猜你喜欢

微积分复习2

About 100 to realize the query table? Really? Let's experience the charm of amiya.

I2C protocol overview

Application of acrelcloud-6000 secure power cloud platform in a commercial plaza

Add sequence number column to MySQL query result set

Introduction to program environment and preprocessing C language (advanced level)

ssh公钥登录失败报错:sign_and_send_pubkey: no mutual signature supported

maya前台渲染插件mel脚本工具

Query the duplicate values of multiple fields in the database, output the number, and add them.

安科瑞抗晃电产品在河北某化工项目的应用
随机推荐
ACL 2022 - strong combination of pre training language model and graphic model
跨域有哪些解决方法?
Selection (045) - what is the output of the following code?
Intel Galileo Gen2 development
Force deduction solution summary 386 dictionary order
Force deduction solution summary 388- longest absolute path of file
Unity3D中DrawCall、Batches、SetPassCall
Force deduction solution summary -04.06 Successor
Acl2022 | DCSR: a sentence aware contrastive learning method for open domain paragraph retrieval
Function templatesfunction templates
Force deduction solution summary 953 verification of alien language dictionary
Apply concentrated load to nodes in batch in ABAQUS
Depth copy
$. map(data,function(item,index){return XXX})
Unique paths for leetcode topic resolution
Getting started with RPC
Force deduction solution summary 905- array sorted by parity
Force deduction solution summary 868- binary spacing
Force deduction solution summary 473 match to square
Solutions to errors in ROM opening by MAME