GenAnalysis: Joint Shape Analysis by Learning Man-Made Shape Generators with Deformation Regularizations 0527 Journal: TOG CCS: Computing methodologies Shape analysis CCS: Neural networks
Yuezhi Yang, Haitao Yang, Kiyohiroe Nakayama, Xiangru Huang, Leonidas Guibas, Qixing Huang · 2025
Read GenAnalysis: Joint Shape Analysis by Learning Man-Made Shape Generators with Deformation Regularizations 0527 Journal: TOG CCS: Computing methodologies Shape analysis CCS: Neural networks by Yuezhi Yang, Haitao Yang, Kiyohiroe Nakayama, Xiangru Huang, Leonidas Guibas, Qixing Huang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-00807.txt.
Introduction
Yuezhi Yang and five co-authors present a shape analysis system in a 2025 paper. Man made objects, like chairs and cars, come in families with shared parts. Software that grasps those parts can match and segment shapes better. Their system, GenAnalysis, learns a generator of such shapes.
A key rule keeps warps between nearby shapes as even as possible. From those warps, the system reads how each shape varies. The variation hints at where part borders lie on one shape. Warps are then chained across many shapes to align them. Shared borders emerge by pooling the single shape hints. Tests on a standard shape library beat older methods, the authors report. The work may aid design tools and 3D search.
Catalog record
| CREATOR | Yuezhi Yang, Haitao Yang, Kiyohiroe Nakayama, Xiangru Huang, Leonidas Guibas, Qixing Huang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 81 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-00807
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
GenAnalysis: Joint Shape Analysis by Learning Man-Made Shape Generators with Deformation Regularizations 0527 Journal: TOG CCS: Computing methodologies Shape analysis CCS: Neural networks Yuezhi Yang, Haitao Yang, Kiyohiroe Nakayama, Xiangru Huang, Leonidas Guibas, Qixing Huang Abstract . We present GenAnalysis, an implicit shape generation framework that allows joint analysis of man-made shapes, including shape matching and joint shape segmentation. The key idea is to enforce an as-affine-as-possible (AAAP) deformation between synthetic shapes of the implicit generator that are close to each other in the latent space, which we achieve by designing a regularization loss. It allows us to understand the shape variation of each shape in the context of neighboring shapes and also offers structure-preserving interpolations between the input shapes. We show how to extract these shape variations by recovering piecewise affine vector fields in the tangent space of each shape. These vector fields provide single-shape segmentation cues. We then derive shape correspondences by iteratively propagating AAAP deformations across a sequence of intermediate shapes. These correspondences are then used to aggregate single-shape segmentation cues into consistent segmentations. We conduct experiments on the ShapeNet dataset to show superior performance in shape matching and joint shape segmentation over previous methods. 1. Introduction Shape analysis is a fundamental research area in geometry processing that enjoys many applications ( Mitra et al. 2014 ; Xu et al. 2017 ) . Existing shape analysis approaches fall into the categories of supervised and unsupervised methods. Although supervised methods achieved state-of-the-art results, they require user labels, which are costly to obtain and do not scale up. In contrast, unsupervised shape analysis methods ( Deng et al. 2021 ; Shuai et al. 2023 ; Chen et al. 2024 ) benefit from learning from large-scale unlabeled shapes and have recently shown promising results in the large-scale ShapeNetPart ( Yi et al. 2016 ) dataset. \begin{overpic}[Figures/teaser.pdf] \put(56.0,-1.0){consistent segmentation} \put(60.0,46.0){shape matching} \put(-3.0,-1.0){as-affine-as-possible deformation} \end{overpic} Figure 1 . GenAnalysis learns a shape manifold from a collection of man-made shapes with a novel as-affine-as-possible deformation regularization loss. The learned shape manifold supports shape analysis by analyzing the tangent space of each shape, shape matching through intermediate shapes on this manifold, and consistent segmentation by aggregating single-shape analysis results using inter-shape correspondences. The fundamental challenge in developing unsupervised techniques is to accommodate significant structural and geometric variations across the input shapes. A popular paradigm is learning template-based models in which each template consists of a bag of (non-rigidly) deforming parts. The learning phase learns parameters of the templates to fit the training data. These include early works of ( Ovsjanikov et al. 2011 ; Kim et al. 2013 ; Averkiou et al. 2014 ; Huang et al. 2015a ; Huang et al. 2015b ) and recent deep learning methods ( Ganapathi-Subramanian et al. 2018 ; Genova et al. 2019a ; Deprelle et al. 2019 ; Shuai et al. 2023 ; Chen et al. 2024 ; Zheng et al. 2021 ; Deng et al. 2021 ) . The key advantage of template-based models is that during the testing phase it is only required to fit template to each shape, which is efficient and simple. However, template-based models exhibit two major limitations. First, template-based models rely on strong prior knowledge of shape parts (e.g., the number of parts and their spatial locations) either explicitly or implicitly. Second, the learning and inference phases can easily get stuck in local minimums. In this paper, we introduce a novel shape analysis framework, named GenAnalysis, which learns an implicit shape generative model to fit the training shapes. We will show how to use this template-free shape generative model to perform shape analysis, hence circumventing the limitation of template-based models. Our approach is inspired by the success of studying organic shape collections using the lens of shape spaces as differential manifolds ( Klassen et al. 2004 ; Kilian et al. 2007 ; Srivastava et al. 2011 ; Hartman et al. 2023 ; Yang et al. 2023a ) . Moreover, recent advances in shape generative models ( Park et al. 2019 ; Chen and Zhang 2019 ; Mescheder et al. 2019 ; Atzmon and Lipman 2021 ; Yang et al. 2024 ) can accurately fit man-made training shapes that exhibit much larger geometric variations than organic shapes. These learned continuous man-made shape spaces offer novel means of shape analysis. Specifically, they provide smooth interpolations between two shapes that can be used to compute shape correspondences. Moreover, when trained properly,these generative models enable us to understand the shape variations among the shape collection for shape segmentation by analyzing shape variations in the tangent space of each shape when viewing the learned generative models as shape manifolds. However, learning a shape generative model is under-constrained, as there are many possible network parameters and latent codes of training shapes that fit the training data. Learning shape generative models under generic distribution alignment paradigms ( Goodfellow et al. 2020 ; Kingma and Welling 2014 ; Zhang et al. 2023 ) do not result in shape generators that offer meaningful intermediate shapes and tangent spaces for shape matching and shape segmentation. Unlike deformable objects in which isometric deformations offer faithful deformation priors, the structural shape variations in man-made shapes are too complex to be summarized into a concise universal principle. The suitable regularization losses are heavily dependent on the application scenarios. GenAnalysis introduces a novel regularization loss for learning the shape generator. This loss is tailored for shape analysis. The proposed loss builds on the popular piece-wise affine assumption ( Xu et al. 2010 ; Ovsjanikov et al. 2011 ; Kim et al. 2013 ; Huang et al. 2015a ; Chen et al. 2024 ) , in which if we abstract each shape part using a bounding box, then the deformations at the part level are affine (see Figure 2 ). It also enforces local piece-wise affine assumption in a more flexible and faithful way compared to template based approaches hence alleviates the local minimum issue by achieving better shape quality. However, the challenge is how to model the significant shape variations within each part. GenAnalysis presents two strategies to address this challenge. First, we minimize the piece-wise affine deformation between adjacent synthesized shapes defined by the generator (i.e., they are close in the latent space). This approach is analogous to the smoothness loss in curve fitting, but we model strong prior knowledge about inter-shape deformations. In particular, we show how to employ an as-affine-as-possible deformation loss under a robust norm to fulfill the piece-wise affine assumption. Second, we describe a weighting scheme for test-time optimization ( Wang et al. 2021 ; Chen et al. 2022 ; Niu et al. 2022b ; Wang et al. 2023 ) . It prioritizes that the regularization loss distributes deviations of piece-wise affine deformations away from shapes of interest. This allows us to extract segmentation cues by analyzing the tangent space of each shape (i.e., vector fields that characterize differences between each shape and its adjacent shapes in the latent space). \begin{overpic}[Figures/PieceWiseAffine.pdf] \put(51.0,28.0){{\color[rgb]{1,0,0}$A_{i}$}} \end{overpic} Figure 2 . Piece-wise affine assumption. Each shape part from the source shape, approximated by its bounding box, undergoes an affine transformation A_{i} to the corresponding part in the target shape. Two chair silhouettes divided into parts;