Where:

Amphi Morse, IMT Nord Europe, Rue Guglielmo Marconi, 59650 Villeneuve-d'Ascq, 

Access with Metro M1, station Cité Scientifique - Professeur Gabillard

When: November 24, 2026, from 9:00 a.m. to 6:00 p.m.

Registration — Free but Mandatory
To ensure access to the workshop room for all participants, registration is free but mandatory.
Participants can attend in person or online. Registered online participants will receive the videoconference link.
👉 Register here (by Novembre 7):  https://evento.renater.fr/survey/second-workshop-of-the-international-research-project-irp-ac8q97xc

9:15 a.m. – Welcome and Presentation of the International Research Project (IRP): Geometric Deep Learning and Generative Models for 3D Humans (Mohamed Daoudi and Stefano Berretti)

9:30 a.m. – 10:30 a.m.

Invited Speaker : Dr. Tolga Birdal, Imperial College London

Title: Generative Neural Priors for 3D Motion Perception

Abstract: Generative modeling has rapidly evolved from a tool for creating synthetic data to a foundational paradigm for solving diverse real-world problems. In this talk, I will present a series of works where generative priors are designed not merely for pose generation, but as robust, geometry-aware motion models that enable reconstruction, recovery, and reasoning under uncertainty. Starting from pose generation via Neural Riemannian Distance Fields (NRDF), I will show how articulated pose distributions (of humans or hands) can be faithfully captured on product manifolds of rotations, yielding powerful priors for inverse kinematics, image-based pose estimation, and denoising. I will then transition to motion models such as HuMoR, which introduces an expressive, dynamics-aware generative model of human motion via conditional VAEs, enabling robust 3D motion recovery even under noise, occlusions, or missing observations. Finally, Neural Riemannian Motion Fields (NRMF) extend these ideas to higher-order motion dynamics, explicitly modeling pose, velocity, and acceleration on motion manifolds through neural distance fields. Together, these models demonstrate how generative formulations serve as flexible priors that power tasks as diverse as denoising, temporal interpolation, spatiotemporal infilling, and recovery of motion from highly incomplete or corrupted data, as evidenced by Dyn-HaMR. By embedding geometry, dynamics, and stochasticity into generative frameworks, the talk will move towards a unified view where generation, perception, and recovery are all facets of the same modeling principle.

Bio: Tolga Birdal is a UKRI Future Leaders Fellow and an associate professor in the Department of Computing of Imperial College London. Previously, he was a Postdoctoral Research Fellow at Stanford University within the Geometric Computing Group of Prof. Leonidas Guibas. Tolga has defended his masters and Ph.D. theses at the Computer Vision Group under Chair for Computer Aided Medical Procedures, Technical University of Munich led by Prof. Nassir Navab. He was also a Doktorand at Siemens AG under supervision of Dr. Slobodan Ilic working on “Geometric Methods for 3D Reconstruction from Large Point Clouds”. His current foci of interest involve topological / geometric machine learning and 3D computer vision. More theoretical work is aimed at investigating and interrogating limits in geometric computing and non-Euclidean inference as well as principles of deep learning. Tolga is an AC for major vision conferences such as CVPR, ICCV, ECCV and 3DV and he has several publications at the well-respected venues such as NeurIPS, CVPR, ICCV, ECCV, ICLR, ICML, T-PAMI, IJCV, ICRA, IROS, ICASSP and 3DV. Aside from his academic life, Tolga has co-founded multiple companies including Befunky, a widely used web-based image editing platform.

Website: tolgabirdal.com

10:30 - 11:00 | Coffe Break

11:00 a.m. – 12:00 a.m.

Invited Speaker : Dr. Gül Varol, Ecole Nationale des Ponts et Chaussées

Title: From Bodies to Hands: Language-Guided 3D Motion Generation

Abstract: This talk will start by reviewing a series of papers covering a variety of tasks bridging natural language and 3D human body motions, such as text-driven motion generation. We will then mainly focus on a more recent work on 3D hand motion generation from language descriptions. In particular, we address the lack of data in this domain by automatically scaling the collection of motion-text pairs. To this end, we introduce (a) BOBSL3DT, which leverages a large-scale sign language video dataset, and (b) HandMDM, which extends motion diffusion model for generating hand pose sequences. We will conclude with a glimpse of more recent works on human-object interaction.

Short bio: Gul Varol is a computer vision researcher in the IMAGINE team at Ecole Nationale des Ponts et Chaussées. Previously, she was a postdoctoral researcher at the University of Oxford (VGG). She obtained her PhD from the WILLOW team of Inria Paris and Ecole Normale Supérieure (ENS). Her thesis received PhD awards from ELLIS and AFRIF. She is the recipient of the ECVA Young Researcher Award 2025. She regularly serves as an Area Chair at major computer vision conferences, and has served as a Program Chair at ECCV'24. Her research interests cover vision and language applications, including video understanding, 3D human motions, and sign languages.

2:00 p.m. – 2:30 p.m.

Speaker: Chaima Zouaoui, IMT Nord Europe

Title: Vertex-VaRefine: Mesh-Agnostic, Pose-Equivariant Refinement of Parametric Human Body Reconstructions.

Abstract: Parametric body models such as SMPL are the standard tool for 3D human reconstruction, but their low-dimensional latent space cannot capture individual body shapes or fine geometric details. We present Vertex-VaRefine, a learning-based method that goes beyond this limit by predicting a per-vertex correction of an existing reconstruction. The gap between the reconstruction and the target is encoded with mesh-agnostic descriptors, so the target can have any topology. A DiffusionNet backbone predicts the correction in rest pose, from descriptors expressed in a local frame attached to each vertex that moves with the body, which makes the model approximately pose-equivariant by construction. On AMASS, Vertex-VaRefine reduces the reconstruction error of the state-of-the-art method ArtEq by about 46%.

2:30 p.m. – 3:00 p.m.

Speaker: Mirko Bicchierai, University of Florence

Title: Beyond Kinematics: Generative Models for Human Body Dynamics

Abstract: Generating realistic digital humans requires more than producing plausible sequences of skeletal poses. Soft-tissue deformations, body morphology, and contact forces all contribute to the physical and perceptual realism of human movement. This talk explores these aspects through two ongoing research directions, with a primary focus on generating soft-tissue dynamics. The first part presents a conditional flow-matching approach that enriches skeletal motion with dynamic soft-tissue deformations learned from registered 4D scans. Deterministic regression can attenuate secondary motion when the available motion information does not fully determine the observed deformations. By treating soft-tissue prediction as conditional generation, the proposed approach produces alternative deformation realizations for the same skeletal motion and captures richer temporal and spectral variation. The discussion examines the balance between positional accuracy and dynamic expressiveness, together with applications to unseen motions and soft-tissue retargeting. The second part introduces a complementary direction: jointly generating human motion and ground-reaction forces conditioned on text and body characteristics. Using a dataset built through musculoskeletal simulation and a diffusion model that generates pose and force sequences together, this work investigates how body morphology and ground interaction can be explicitly incorporated into motion generation.

3:00 p.m. – 4:00 p.m

Invited Speaker:  Dr. Eric Paquet, Senior Research Officer at the National Research Council Canada

Tittle:  Conditioned quantum-assisted generative simulation Calo4pQVAE: a surrogate for particle–calorimeter interactions

Abstract: Detailed calorimeter simulations are essential in high-energy physics but are computationally expensive. We present Calo4pQVAE, a hybrid quantum-classical generative model designed to produce realistic calorimeter showers conditioned on the incident particle energy. The model combines a discrete variational autoencoder with a four-partite restricted Boltzmann machine whose structure is tailored to the connectivity of a D-Wave quantum annealer. Classical Gibbs sampling and quantum annealing are used as alternative mechanisms for sampling the learned latent distribution. The work also introduces a practical conditioning scheme based on flux biases and an adaptive calibration of the QPU effective temperature. Results show that quantum-generated latent samples can reproduce key calorimeter observables comparably to classical sampling, while highlighting the remaining challenges associated with conditioning overhead and end-to-end performance.

4:00 - 04:30 | Coffe Break

4:30 p.m. – 5:00 p.m.

Speaker: Thierno Fall, University Iba Der Thiam, Thies 

Title: Equivariant Generative Models Based on Riemannian Morphological Operators: Application to GANs and DDPMs.

Abstract: Standard generative models rely on convolutional architectures that are only translation-equivariant, which limits their ability to respect other geometric symmetries of the data. In this talk, we show how equivariant inductive biases can be built into generative models through layers derived from morphological PDEs. We first present GM-GAN, which integrates equivariant PDE-G-CNN layers into GANs, improving sample quality and robustness to geometric transformations. We then extend this framework to diffusion models with REMoD, whose noise-prediction network combines group morphological convolutions and a convection operator defined on the sphere. Experiments show improved quality of the generated images and, on rotated data, the stability of these models.

5:00 pm - 6:00 pm

Speaker: Dr. Claudio Ferrari, University of Florence

Title: 3D Face Modeling and Animation: Current Prospects and Challenges

Abstract: Representing, modeling and animating 3D human faces is a long standing problem in computer vision and machine learning, with a broad range of downstream applications spanning from entertainment, graphics and biometrics to the medical field. Despite being studied for decades, the complexity of the problem still challenges researchers, and the room for improvement is large. A lot is yet to be done. In this talk, we’ll delve into the current solutions, challenges and prospects of the task of reconstructing, modeling and animating 3D faces. The talk will introduce state-of-the-art techniques for representing and modeling the geometry of faces, and how such techniques were leveraged to develop solutions for 3D/4D animation, with a specific focus on facial expressions and speech generation (Talking Heads). From a different perspective, the talk will also address another thorny problem: how can we accurately measure progress? While properly measuring error is of paramount importance for accurate assessment of a method performance, current options face several limitations and fall short under different perspectives.