This website requires JavaScript.

Neural Shape Compiler: A Unified Framework for Transforming between Text, Point Cloud, and Program

Tiange LuoHonglak LeeJustin Johnson
Dec 2022
摘要
3D shapes have complementary abstractions from low-level geometry topart-based hierarchies to languages, which convey different levels ofinformation. This paper presents a unified framework to translate between pairsof shape abstractions: $\textit{Text}$ $\Longleftrightarrow$ $\textit{PointCloud}$ $\Longleftrightarrow$ $\textit{Program}$. We propose $\textbf{NeuralShape Compiler}$ to model the abstraction transformation as a conditionalgeneration process. It converts 3D shapes of three abstract types into unifieddiscrete shape code, transforms each shape code into code of other abstracttypes through the proposed $\textit{ShapeCode Transformer}$, and decodes themto output the target shape abstraction. Point Cloud code is obtained in aclass-agnostic way by the proposed $\textit{Point}$VQVAE. On Text2Shape,ShapeGlot, ABO, Genre, and Program Synthetic datasets, Neural Shape Compilershows strengths in $\textit{Text}$ $\Longrightarrow$ $\textit{Point Cloud}$,$\textit{Point Cloud}$ $\Longrightarrow$ $\textit{Text}$, $\textit{PointCloud}$ $\Longrightarrow$ $\textit{Program}$, and Point Cloud Completion tasks.Additionally, Neural Shape Compiler benefits from jointly training on allheterogeneous data and tasks.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?