Geometric Learning and Facial Animation

My work in geometric learning uses geometric priors to construct models for 3D data. SCENE-Net studies interpretable 3D segmentation through Group Equivariant Non-Expansive Operators. Related work addresses shape modelling from incomplete observations.

In facial animation, I study optimisation methods for inverse rigging and the estimation of blendshape controls. This connects the geometry of a facial model with the computation needed to produce animation.

Related work examines the geometry of vision-language representations through spectral diagnostics, including evaluation on medical image-text retrieval. The 3D perception work also addresses infrastructure inspection using point clouds of electrical transmission systems.

Facial-motion generation

STEmo is a research dataset described in a manuscript under review. It records facial motion during transitions between emotion categories, with generation baselines used to study how these transitions are reproduced.

This line of work also examines how facial-motion generation and preview can be integrated into Unreal Engine workflows for MetaHuman characters.

Funded projects

Projects active during 2021 or later, including completed projects.

• BIGMATH (2018–2022). National principal investigator. Industrial doctoral research in optimisation, statistics and linear algebra, including facial data analysis. The University of Milan coordinated the consortium. Reference 812912.

Publications since 2021

• SCENE-Net: Geometric induction for interpretable and low-resource 3D pole detection with Group-Equivariant Non-Expansive Operators. Computer Vision and Image Understanding, 2025.
• Learning Under Noisy Labels, Spurious Points, and Diverse Structures: TS40K, a 3D Point Cloud Dataset of Rural Terrain and Electrical Transmission Systems. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV).
• Refined Inverse Rigging: A Balanced Approach to High-fidelity Blendshape Animation. SIGGRAPH Asia 2024 Conference Papers.
• SCENE-Net V2: Interpretable Multiclass 3D Scene Understanding with Geometric Priors. ICML 2024 Workshop GRaM.
• Distributed Solution of the Blendshape Rig Inversion Problem. SIGGRAPH Asia 2023 Technical Communications.
• Probabilistic Registration for Gaussian Process Three-Dimensional Shape Modelling in the Presence of Extensive Missing Data. SIAM Journal on Mathematics of Data Science, 2023.
• A majorization-minimization-based method for nonconvex inverse rig problems in facial animation: algorithm derivation. Optimization Letters, 2023.
• Clustering of the Blendshape Facial Model. 2021 29th European Signal Processing Conference (EUSIPCO).
• From Noisy Point Clouds to Complete Ear Shapes: Unsupervised Pipeline. IEEE Access, 2021.

Preprints and related manuscripts since 2021

• Beyond Symmetric Alignment: Spectral Diagnostics of Modality Imbalance in Vision-Language Models in the Medical Domain. arXiv preprint, 2026.
• Enhancing Power Grid Inspections with Machine Learning. arXiv preprint, 2025.
• Low-Resource White-Box Semantic Segmentation of Supporting Towers on 3D Point Clouds via Signature Shape Identification. arXiv preprint, 2023.
• High-fidelity Interpretable Inverse Rig: An Accurate and Sparse Solution Optimizing the Quartic Blendshape Model. arXiv preprint, 2023.
• Accurate and Interpretable Solution of the Inverse Rig for Realistic Blendshape Models with Quadratic Corrective Terms. arXiv preprint, 2023.
• Accurate, Interpretable, and Fast Animation: An Iterative, Sparse, and Nonconvex Approach. arXiv preprint, 2021.