Mesh-based Autoencoders for Localized Deformation Component Anal...
198 浏览量
2021-02-07
16:54:57
上传
评论
收藏 1.11MB PDF 举报
Mesh-based Autoencoders for Localized Deformation Component Analysis
Qingyang Tan
1,2
, Lin Gao
1∗
, Yu-Kun Lai
3
, Jie Yang
1,2
and Shihong Xia
1
1
Beijing Key Laboratory of Mobile Computing and Pervasive Device,
Institute of Computing Technology, Chinese Academy of Sciences
2
School of Computer and Control Engineering, University of Chinese Academy of Sciences
3
School of Computer Science & Informatics, Cardiff University
tanqingyang14@mails.ucas.ac.cn, {gaolin, yangjie01, xsh}@ict.ac.cn, LaiY4@cardiff.ac.uk
Abstract
Spatially localized deformation components are very useful
for shape analysis and synthesis in 3D geometry processing.
Several methods have recently been developed, with an aim to
extract intuitive and interpretable deformation components.
However, these techniques suffer from fundamental limita-
tions especially for meshes with noise or large-scale defor-
mations, and may not always be able to identify important
deformation components. In this paper we propose a novel
mesh-based autoencoder architecture that is able to cope with
meshes with irregular topology. We introduce sparse regular-
ization in this framework, which along with convolutional op-
erations, helps localize deformations. Our framework is ca-
pable of extracting localized deformation components from
mesh data sets with large-scale deformations and is robust to
noise. It also provides a nonlinear approach to reconstruction
of meshes using the extracted basis, which is more effective
than the current linear combination approach. Extensive ex-
periments show that our method outperforms state-of-the-art
methods in both qualitative and quantitative evaluations.
1 Introduction
With the development of 3D scanning and modeling tech-
nology, mesh data sets are becoming more and more popu-
lar. By analyzing these data sets with machine learning tech-
niques, the latent knowledge can be exploited to advance
geometry processing algorithms. In recent years, many re-
search areas in geometry processing have benefited from
this methodology, such as 3D shape deformation (Gao et
al. 2016), 3D facial and human body reconstruction (Cao et
al. 2015; Bogo et al. 2016), shape segmentation (Guo, Zou,
and Chen 2015), etc. For shape deformation and human re-
construction, mesh sequences with different geometry and
the same connectivity play a central role. Different geomet-
ric positions describe the appearance of the 3D mesh model
while sharing the same vertex connectivity makes process-
ing much more convenient. In such works, a key procedure
is to build a low-dimensional control parametrization for the
mesh data set, which provides a small set of intuitive param-
eters to control the generation of new shapes. For articulated
models such as human bodies, the rigging method embeds a
∗
Corresponding Author
Copyright
c
2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
skeleton structure in the mesh to provide such a parametriza-
tion. However, the rigging operation is restrictive and does
not generalize to other deformable shapes (e.g. faces). Pa-
rameterizing general mesh datasets which allows intuitive
control in generating new shapes becomes an important and
urgent research problem.
Early work extracted principal deformation components
by using Principal Component Analysis (PCA) to reduce
the dimensionality of the data set. However, such defor-
mation components are global which do not lead to intu-
itive control. For example, when a user intends to deform
the shape locally by specifying locally changed vertex po-
sitions as boundary conditions, the deformed shape tends
to have unrelated areas deformed as well, due to the global
nature of the basis. To address this, sparse localized defor-
mation component (SPLOCS) extraction methods were re-
cently proposed (Neumann et al. 2013; Huang et al. 2014;
Wang et al. 2016). In these works the sparsity term is in-
volved to localize deformation components within local sup-
port regions. However, these previous works suffer from dif-
ferent limitations: as we will show later, (Neumann et al.
2013; Huang et al. 2014) cannot handle large-scale defor-
mations, and (Wang et al. 2016) is sensitive to noise which
cannot extract the main deformation components robustly.
We propose a novel mesh-based autoencoder architecture to
extract meaningful local deformation components. We rep-
resent deformations of shapes in the dataset based on a re-
cent effective representation (Gao et al. 2017) which is able
to cope with large deformations. We then build a CNN-based
autoencoder to transform the deformation representation to
encoding in a latent space. Each convolutional layer involves
convolutional operations defined on the mesh with arbitrary
topology in the form of applying the same local filter to
each vertex and its 1-ring neighbors, similar to (Duvenaud
et al. 2015). We then introduce sparsity regularization to the
weights in the fully-connected layers to promote identify-
ing sparse localized deformations. The autoencoder struc-
ture ensures that the extracted deformation components are
suitable for reconstructing high quality shape deformations.
Our main contributions are: 1) This is the first work that
exploits CNN-based autoencoders for processing meshes
with irregular connectivity. 2) Benefiting from sparse regu-
larization and the nonlinear representation capability of au-
toencoders, our method is able to extract intuitive localized
PRELIMINARY VERSION: DO NOT CITE
The AAAI Digital Library will contain the published
version some time after the conference
资源评论
