Talks and Abstracts

- Title: 3D analysis-by-synthesis

- Abstract: In this talk, I will discuss how the analysis-by-synthesis paradigm can be applied to 3D data. In a first part, I will introduce a way to model surfaces using a deep deformation network [1] and show how it can be used to model correspondences [2], and discover object parts [3] and scene components [4] in an unsupervised way. In a second part, I will discuss how such models can be included in an image rendering pipeline to infer 3D scene geometry from images. In particular, I will discuss how this can be used to learn single view reconstruction without any supervision [5] and to discover geometric primitives in mutli-view images [6].

- Bio: Mathieu Aubry is a tenured researcher in Computer Vision at École des Ponts ParisTech in the LIGM lab (UMR8049).  He obtained his PhD at ENS in 2015, co-advised by Josef Sivic (INRIA) and Daniel Cremers (TUM). In 2015, he spent a year working as a postdoc with Alexei Efros in UC Berkeley. He has a leading role in the ANR EnHerit, VHS and EIDA projects and the ERC DISCOVER project on interpretable visual structures discovery.

- Title: Is Human Motion a Language without Words?

- Abstract: This talk will summarize our recent works on bridging the gap between natural language and 3D human motions. I will first show results on text-to-motion synthesis, i.e., text-conditioned generative models for controllable motion synthesis, with a special focus on compositionality to handle finegrained textual descriptions. Second, I will present results from our text-to-motion retrieval model. The relevant papers are ACTOR, TEMOS, TMR [Petrovich 2021, 2022, 2023] and TEACH, SINC [Athanasiou 2022, 2023]

- Bio: Gül Varol is a permanent researcher in the IMAGINE team at École des Ponts ParisTech. Previously, she was a postdoctoral researcher at the University of Oxford (VGG). She obtained her PhD from the WILLOW team of Inria Paris and École Normale Supérieure (ENS). Her thesis received the PhD awards from ELLIS and AFRIF. Her research is focused on computer vision, specifically video representation learning, human motion analysis, and sign languages.

- Title: Learning Geometric Models from Limited Data

- Abstract: 3D Content Creation remains difficult and expensive. While large-scale generative models have rapidly emerged in the image world, a similar success story still needs to be discovered in the 3D world. What will it take to make this happen? In this talk, I will discuss the challenges with representations and the lack of direct training data. Without direct supervisory data, one option is to train with indirect signals via an image-formation model or indirectly through observed (character) motion. I would report our current findings and show initial results that indicate that one may be able to tackle this problem using indirect supervision, often in the form of geometry-guided inductive bias. I will demonstrate applications in 3D modeling and character motion.

- Bio: Niloy J. Mitra leads the Smart Geometry Processing group in the Department of Computer Science at University College London and the Adobe Research London Lab. He received his Ph.D. from Stanford University under the guidance of Leonidas Guibas. His research focuses on developing machine learning frameworks for generative models for high-quality geometric and appearance content for CG applications. He was awarded the Eurographics Outstanding Technical Contributions Award in 2019, the British Computer Society Roger Needham Award in 2015, and the ACM SIGGRAPH Significant New Researcher Award in 2013. Furthermore, he was elected as a fellow of Eurographics in 2021 and served as the Technical Papers Chair for SIGGRAPH in 2022. His work has also earned him a place in the SIGGRAPH Academy in 2023. Besides research, Niloy is an active DIYer and loves reading, cricket, and cooking. For more, please visit his webpage..

- Title: Supervised ultrametric learning for hierarchical segmentation

- Abstract: Hierarchical image segmentation aims to capture the structure of objects of different sizes at different scales in a scene. With the success of neural networks for image segmentation, the task of learning of segmentation hierarchies naturally arises. However, it remains difficult to impose such a structural constraint on the output of neural networks. In this talk, we formulate this as the problem of learning an ultrametric distance, which is a functional representation of a hierarchical segmentation. We introduce an ultrametric layer that can transform any dissimilarity function into an ultrametric distance. We discuss possible hierarchical loss functions and show how this new layer can be used to train neural networks to predict hierarchical segmentation.

- Bio: Benjamin Perret received his Ph.D. in Image Analysis from the Université de Strasbourg, France in 2007 and his Habilitation from the Université Paris-Est in 2021. He is currently a professor at ESIEE Paris, Université Gustave Eiffel, affiliated with the Laboratoire d'Informatique Gaspard Monge. His current research interests include image analysis, machine learning, and mathematical morphology.

- Title: Some Geometry of Feature Learning for Non-Rigid 3D Shapes

- Abstract: Learning informative pointwise features is a key task in 3D shape analysis with broad applications in many downstream applications. In this talk, I will present several recent approaches for feature learning, and emphasize the influence of geometric considerations during this feature learning process. I will first describe a link between contrastive learning and non-rigid shape correspondence, informing both of these areas. I will then mention how robust and generalizable features can be trained by connecting rigid and non-rigid shape analysis. Finally, I will talk about how unsupervised feature learning can give rise to surprisingly informative and rich features that can be exploited for other problems. Throughout my talk, I will emphasize especially how learning can be done in limited data (or annotation) regimes.

- Bio: Maks Ovsjanikov is a Professor at Ecole Polytechnique in France. He works on 3D shape analysis with emphasis on deep learning techniques for shape matching and correspondence. He has received a Eurographics Young Researcher Award, an ERC Starting Grant, a CNRS Bronze Medal, and an ERC Consolidator Grant in 2023. His works have received 11 best paper awards or nominations at top conferences, including CVPR, ICCV, 3DV, etc. His main research topics include 3D shape comparison and deep learning on 3D data.

- Title: Topological Machine Learning with applications in Computational Biology

- Abstract:Topological Data Analysis (TDA) is a field of data science that is gathering increasing attention in the recent years, due to its ability to produce descriptors of topological flavor that (a) can be computed on a wide range of data sets, and (b) encode a unique type of information that is very useful yet missed by other standard descriptors. However, integration of such descriptors in standard machine learning pipelines is not straightforward, be it Mapper complexes or persistent homology. In this talk, I will present the main difficulties associated to the creation of topological machine learning pipelines, and then I will go over the various methods and solutions proposed over the recent years to handle these issues, including, e.g., the statistical treatment of Mapper, the representation of (multi-parameter) persistent homology, and the differentiability of topological descriptors. I will also illustrate these methods on a few applications from computational biology.

- Bio: Mathieu did his PhD at Inria Saclay in the DataShape team, under the supervision of Steve Oudot, and a postdoc of two years in the Rabadán Lab, at the Department of Systems Biology of Columbia University, under the supervision of Raúl Rabadán. His research focuses on topological data analysis and statistical machine learning, with an application to bioinformatics and genomics. He contributed to the analysis of topological descriptors and their use in machine learning tools such as kernel methods and deep learning.

- Title: From morphological-aided deep learning to learning morphological models from data

- Abstract: This talk presents recent work that have been done by the morphological community on the interactions between Mathematical Morphology (MM) and Deep Learning (DL). First, I will give an overview of the powerful and elegant features that MM provides for the topological analysis of images. Secondly, I will show how these can be used efficiently to improve machine learning approaches, either through pre/post processing or (pseudo)labeling. Finally, the recent progress and main current issues of morphological neural networks will be presented.

- Bio: Samy graduated from Telecom Paris and the MVA masters degree (Mathématiques Visions Apprentissage) and did his PhD at ENS Paris-Saclay, on the comparison of a probabilistic approach of image analysis (called "a contrario" methods) to human vision. Since 2017 he has been working at the Centre for Mathematical Morphology (CMM) of Mines Paris, on theoretical and practical developments in both mathematical morphology and deep learning for image processing, as well as their interactions..