DSM-Net: Disentangled Structured Mesh Net for Controllable Generation of Fine Geometry
Dsm-net: Disentangled structured mesh net for controllable generation of fine geometry
目录
Motivation
- 把structure(topology)和geometry进一步解耦,in a synergistic manner
Overview
- 用Recursive Neural Networks(RvNNs, 注意RNN是recurrent NN) hierarchically encode和decode structure和geometry,在hierarchy的每一层都有bijective mapping
- 同时用两个分开的但是高度耦合的VAE学习structure 和geometry,把他们encode into two latent spaces
disentangled shape representation
structure hierarchy抽象出符号部件(symbolic parts)与关系
- inspired by PT2PC: Learning to Generate 3D Point Cloud Shapes from Part Tree Conditions. 2020
- 每个部件用semantic label (e.g. chair back, chair leg)表示,引入PartNet dataset中丰富的部件关系
- $\boldsymbol{\rm H}$ 纵向的parent-child inclusion 关系 (e.g. chair back and chair back bars)
- $\boldsymbol{\rm R}$ 横向的among-sibling 部件对称性与邻接性(e.g. chair back bars have translational symmetry)
geometry hierarchy是部件的geometry
- 表征就是正常的多顶点mesh
- 假设一个5402顶点构成的封闭mesh,计算oriented bounding box
- 然后通过non-rigid registration 变形这个mesh到target part geometry
- 然后用ACAP作为部件表征
- Sparse data driven mesh deformation. 2019
- SDM-NET: Deep Generative Network for Structured Deformable Mesh. 2019
structure hierarchy和geometry hierarchy之间有bijective mapping
- 符号部件 $l_i$ 对应部件geometry $G_i$ ,层级 $\boldsymbol{\rm H}$ 和关系 $\boldsymbol{\rm R}$ 则隐式地互相一致
- 在学习的时候两个hierarchies有communication channels
- 虽然结构和几何要解耦,但是他们还是需要彼此兼容来产生好的、现实的形状
- 一方面,shape structure 为 part geometry提供high-level guidance
- e.g. 如果four legs of a chair对称,那么他们应该具有identical part geometry
- 另一方面,给定part geometry以后,只有若干种适用的shape structures(而不是全部)
- e.g. 如果没有lift handle或者gas cylinder parts,不可能组装一个swivel chair
- 一方面,shape structure 为 part geometry提供high-level guidance
conditional part geometry VAE
- encode和decode时候都condition on part structure information
Disentangled Geometry and Structure VAEs
- 下图蓝色代表geometry,红色代表structure
encoding的时候,从geometry和structure feature encode出geometry
decoding的时候,从geometry和structure feature decode出geometry - Q: what?