CoChat

Learning skeletal articulations with neural blend shapes

Li, Peizhuo, Aberman, Kfir, Hanocka, Rana, Liu, Libin, Sorkine-Hornung, Olga, Chen, Baoquan · 2021 · ACM Transactions on Graphics

The DOI matches its Crossref record; 66 of the 66 references it cites were found in a registry or an index; 5 of the 16 sources its claims cite were checked for support.

Abstract

Animating a newly designed character using motion capture (mocap) data is a long standing problem in computer animation. A key consideration is the skeletal structure that should correspond to the available mocap data, and the shape deformation in the joint regions, which often requires a tailored, pose-specific refinement. In this work, we develop a neural technique for articulating 3D characters using enveloping with a pre-defined skeletal structure which produces high quality pose dependent deformations. Our framework learns to rig and skin characters with the same articulation structure ( e.g. , bipeds or quadrupeds), and builds the desired skeleton hierarchy into the network architecture. Furthermore , we propose neural blend shapes - a set of corrective pose-dependent shapes which improve the deformation quality in the joint regions in order to address the notorious artifacts resulting from standard rigging and skinning. Our system estimates neural blend shapes for input meshes with arbitrary connectivity, as well as weighting coefficients which are conditioned on the input joint rotations. Unlike recent deep learning techniques which supervise the network with ground-truth rigging and skinning parameters, our approach does not assume that the training data has a specific underlying deformation model. Instead, during training, the network observes deformed shapes and learns to infer the corresponding rig, skin and blend shapes using indirect supervision. During inference, we demonstrate that our network generalizes to unseen characters with arbitrary mesh connectivity, including unrigged characters built by 3D artists. Conforming to standard skeletal animation models enables direct plug-and-play in standard animation software, as well as game engines.

Overview

CoChat Summary

TL;DR

The paper presents a neural framework that rigs and skins a character mesh with a prescribed skeleton hierarchy and predicts neural blend shapes for pose-dependent corrective deformation, using indirect supervision from articulated vertex positions. It reports improved rigging and skinning accuracy over Pinocchio and RigNet and demonstrates generalization to unseen characters with arbitrary mesh connectivity.

Key findings

  • Rigging accuracy: Table 1 reports CD-J2J / CD-J2B / CD-B2B of 0.012 / 0.007 / 0.004 for the proposed method, versus 0.194 / 0.084 / 0.009 for RigNet and 0.474 / 0.164 / 0.025 for Pinocchio.
  • Skinning and deformation: Table 2 reports skinning weight L1 error 2.56 for the proposed method versus 23.1 for Pinocchio; average distance 0.011 versus 0.27; max distance 1.68 versus 20.3.
  • Ablation: Table 3 reports L2 error 0.011 with all components, 0.024 with envelope only, 0.66 without mesh augmentation, and 0.23 without garment augmentation; the paper states mesh augmentation has the most quantitative impact, while neural blend shapes are extremely important for visual quality.
  • The paper reports that 9 blend shapes were enough for high-quality pose-dependent deformations, and that indirect supervision converges slower than direct supervision (80,000 vs. 45,000 iterations to reach Lv = 0.024) but with comparable deformation quality.

Methods

  • The framework has two branches: an envelope deformation branch that predicts skeletal offsets O and skinning weights W using MeshCNN and skeleton-aware operators, and a residual deformation branch that predicts N blend shapes plus pose-dependent blending coefficients conditioned on joint rotations, with a mask for localization.
  • Training uses the SMPL dataset (10 shape blend shapes and 207 pose-dependent blend shapes) and the Multi Garment Network dataset (96 clothed humans; 80 for training, 16 for testing), with mesh connectivity augmentation (edge collapse, split, flip) and garment augmentation. Test groups are 5 animator-created biped characters, 30 random SMPL characters, and 16 held-out clothed characters; Mixamo characters are used to test generalization.
  • Training is implemented in PyTorch with a two-stage coarse-to-fine approach and Adam, taking about 3 days; supervision is indirect, using an L2 loss on articulated vertex positions rather than ground-truth rigging or skinning parameters.

Limitations

  • Stated: the output is bounded by the quality of deformations in the training dataset; most results are trained using SMPL-like models, and artifacts can appear where the selected model was not trained sufficiently. Changing the prescribed skeleton structure requires retraining the entire network.
  • Stated: the network assumes a consistent upright and front-facing orientation and a T-pose input; non-manifold edges or vertices should be deleted, and extreme polygon soup may require remeshing or tetrahedralization.
  • Evident from the text: the system is trained on characters with the same articulation structure (e.g., bipeds) and evaluated on human/humanoid characters, so generalization to other articulation structures or arbitrary skeleton hierarchies is not quantitatively demonstrated.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3450626.3459852

Cites

66 of 66 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

3 Valid · 3 Not valid · 7 Not checked · 64 Own results and interpretations

37 more claims

Create a free account to see all 77 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 30
  • Probably not a substantial claim. 7
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Citations

Cites

Research works printed but not linked: 3

  1. 10.1111/j.1467-8659.2012.03034.x_eprint · doi:10.1111/j.1467-8659.2012.03034.x_eprint
  2. Jim Hejl . 2004 . Hardware skinning with quaternions . Game Programming Gems 4 (2004), 487 -- 495 . Jim Hejl. 2004. Hardware skinning with quaternions. Game Programming Gems 4 (2004), 487--495.
  3. Tianxing Li , Rui Shi , and Takashi Kanai . 2020 . DenseGATs: A Graph-Attention-Based Network for Nonlinear Character Deformation. In Symposium on Interactive 3D Graphics and Games. 1--9. Tianxing Li, Rui Shi, and Takashi Kanai. 2020. DenseGATs: A Graph-Attention-Based Network for Nonlinear Character Deformation. In Symposium on Interactive 3D Graphics and Games. 1--9.

Cited by

See all 6

Metadata

Authors
Li, P., Aberman, K., Hanocka, R., Liu, L., Sorkine-Hornung, O., & Chen, B.
Affiliations
  • Peking University (CN)
  • Google (United States) (US)
  • Tel Aviv University (IL)
  • ETH Zurich
Venue
ACM Transactions on Graphics
Published
2021-07-19
Volume
40
Issue
4
Pages
1-15
ISSN
  • 0730-0301 (ISSN-L)
  • 1557-7368
OpenAlex
W4235221956
Cited by
4
References
47
Citations by year
2024: 2 · 2023: 1 · 2022: 1
FWCI
0.57
Type
journal article
Language
en
Open access
closed
Locations
ACM Transactions on Graphics (publishedVersion)
Topics
  • Physical Sciences › Computer Science › Computer Graphics and Computer-Aided Design › 3D Shape Modeling and Analysis
  • Physical Sciences › Computer Science › Computer Graphics and Computer-Aided Design › Human Motion and Animation
  • Physical Sciences › Computer Science › Computer Graphics and Computer-Aided Design › Computer Graphics and Visualization Techniques
Keywords
skeletal animation, character deformation, artificial neural networks

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.