Seminars – 2021-2022

The seminars of OBELIX team are currently held on thursdays, every two weeks, at the IRISA lab, Tohannic campus.

The seminar is coordinated by Chloé Friguet.

Previous seminars – 2021-2022


  • Date:  June, 13th (Monday)
  • Time: 11h30
  • Room : D001 (bat. ENSIBS)
  • Speaker: Corentin Dufourg (IRISA-OBELIX)
  • Title:  Étude et conception de réseaux de neurones pour le suivi des radeaux de sargasses à partir d’images satellites
  • Abstract: Depuis plusieurs années, les Caraïbes sont envahies par les sargasses, des algues brunes pélagiques impactant fortement les populations locales. Dans le but d’analyser et prévoir ces échouements, ce travail se concentre sur la télédétection en haute mer des radeaux d’algues grâce aux images des satellites Sentinel-3. Ce suivi des algues passe en partie par un problème de segmentation d’instance cherchant à relever les défis offerts par la résolution limitée des images satellites et par les objets déformables et sécables que sont les radeaux d’algues. Dans cette présentation, nous allons explorer une approche utilisant la segmentation sémantique suivie d’un partitionnement de graphe par réseau de neurones afin de réaliser une segmentation d’instance d’objets fins et entremêlés.

  • Date:  May, 12th
  • Time: 13h45
  • Room : Amphi 102 (bat. DESG)
  • Speaker: Laetitia CHAPEL (IRISA-OBELIX) HDR Defense
  • Title:  Machine learning for structured data

  • Date:  March, 31st
  • Time: 11:30
  • Room : D-001 (bat ENSIBS)
  • Speaker: Adam Herout (Brno University of Technology, Department of Computer Graphics and Multimedia, Rep. Tcheque)
  • Title:  Self-Supervised Learning for Sports Pose Recognition/Classification 
  • Abstract: We intend to help athletes with correct practice in sports poses. For that, we are developing algorithms of efficient/cheap classification of sports poses in real time on mobile devices. The poses should be learned in a few-shot manner. This leads to exploration of self-supervised learning methods and specific datasets and their processing. The talk might be interesting not only for someone interested in sports data processing, but for anyone interested in deep learning, few-shot learning and self-supervised learning.

  • Date:  March, 23th (wednesday)
  • Time: 14h00
  • Room : Paris (ONERA) and Visio
  • Speaker: Javiera CASTILLO ( PhD student, ONERA + IRISA) PhD Defense
  • Title:  Semi-supervised learning for large-scale Earth observation data understanding.
  • Abstract: Earth observation (EO) plays a significant role in the way we understand our planet and its dynamics. While plenty of data are available, they cannot be processed by humans only, so artificial intelligence has emerged as a solution to achieve automatic analysis of EO imagery. Still, most data are not exploited because they are unlabeled. Hence, algorithms beyond supervised learning are needed to get complete insight.
    This thesis investigates deep semi-supervised learning (SSL) for classification and segmentation in order to achieve EO data understanding at a large scale. First, we explore the potential of unlabeled data and propose tools for analyzing data representativeness for multi-location datasets. Then, we explore two ways of approaching the SSL problem. By discriminative modeling, first, we develop multi-task networks and auxiliary tasks to tackle semi-supervised semantic segmentation; second, we explore consistency regularization methods (e.g., FixMatch) to perform scene classification in EO data. Moving to generative modeling, we show the potential of joint energy-based models for semi-supervised classification and many other EO applications.
    Through extensive experiments, we show that SSL allows us to train algorithms with better performances and generalization capacities for land use and land cover mapping.
    Finally, our contributions also include the release of MiniFrance, the first dataset and open benchmark designed to assess and help design SSL in remote sensing, and part of the IEEE GRSS Data Fusion Contest 2022.

  • Date:  March, 10th
  • Time: 11:30
  • Room : D-001 (bat ENSIBS)
  • Speaker: Raquel Almeida (PhD. Student, IRISA-Linkmedia)
  • Title: Learning hierarchies: challenges and what the data tell us
  • Abstract: Hierarchies are rich structures that contain in itself broad information about the data and  their relationship. With the advent of deep learning and more elaborated machine learning  algorithms, the interest in learning hierarchies to create more succinct representations that remain  informative grew. Despite the interest and multiple possible applications, learning hierarchies pose many challenges from their structure, content, and size. Despite these challenges, we can investigate the data itself and the knowledge we can extract from inspectable algorithms that can give clues on how to summarize and filter hierarchies to apply it to machine learning algorithms. In this presentation, we will explore the most known hierarchies, expose the main challenges of learning hierarchies, detail the advances made on understanding the data and how it can be utilized to create more manageable structures for machine learning.

  • Date:  december, 14th
  • Time: 9:00
  • Room: IRISA Rennes Salle Métivier (and visio)
  • Speaker: Heng Zhang (PhD defense)
  • Title: Multispectral Object Detection
  • Abstract:  Scene analysis with only visible cameras is challenging when facing insufficient illumination or adverse weather. To improve the recognition reliability, multispectral systems add additional cameras (e.g. infra-red) and perform object detection from multispectral data. Although the concept of multispectral scene  analysis with deep learning has great potential, there are still many open research questions and it has not been  widely deployed in industrial contexts. In this thesis, we investigated three main challenges about multispectral  object detection: (1) the fast and accurate detection of objects of interest from images; (2) the dynamic and  daptive fusion of information from different modalities; (3) low-cost and low-energy multispectral object detection and the reduction of its manual annotation efforts. In terms of the first challenge, we first optimize the label assignment of the object detection training with a mutual guidance strategy between the classification and localization tasks; we then realize an efficient compression of object detection models by including the teacher-student prediction disagreements in a feature-based knowledge distillation framework. With regard to the second challenge, three different multispectral feature fusion schemes are proposed to deal with the most difficult fusion cases where different cameras provide  contradictory information. For the third challenge, a nouvel modality distillation framework is firstly presented to tackle the hardware and software constraints of current multispectral systems; then a multi-sensor-based active learning strategy is designed to reduce the labelling costs when constructing multispectral datasets.
    Jury : Vincent LePetit / Tinne Tuytelaars / Patrick Bouthemy / Patrick Perez /Jakob Verbeek/ Elisa Fromont / Sébastien Lefèvre

  • Date:  december, 9th
  • Time : 11h30
  • Room: VISIO
  • Speaker: Thibault Séjourné (PhD candidate, ENS Paris)
  • Title: Distance de Gromov-Wasserstein ‘unbalanced’
  • Abstract: Le transport optimal, théorie permettant de définir des distances entre distributions, est un outil de choix dans les domaines de l’apprentissage machine et de l’estimation statistique car il prend en compte la géométrie de l’espace sous-jacent. Ces distances souffrent cependant de trois limitations pouvant être problématiques : (i) elles sont coûteuses à calculer, (ii) se limitent à la comparaison de probabilités et (iii) comparent des mesures définies sur le même espace. Ces contraintes peuvent être gênantes pour passer à l’échelle dans les calculs, pour être insensible aux “outliers” géométriques (dûs à des données bruitées), ou comparer des graphes (tels que des molécules aux structures différentes). Pour pallier ces limitations ont été proposés la régularisation entropique, le transport non-équilibré et la distance de Gromov-Wasserstein. Dans cette présentation, j’introduirai d’abord la formulation non-équilibrée du transport, ainsi que sa variante entropique. Je détaillerai une variante de l’algorithme de Sinkhorn permettant de calculer le dual du problème grâce à une modification mineure de l’algorithme dans sa version équilibrée, avec une convergence linéaire. Dans un second temps, je présenterai la distance de Gromov-Wasserstein qui est un problème d’optimisation quadratique non convexe comparant des espaces munis d’une métrique et d’une mesure positive. Je définirai deux généralisations non-équilibrée de cette distance, l’une étant une borne supérieure pour l’autre. Je montrerai que la première définit une distance entre espaces métriques mesurés, et que pour la seconde il est possible de la calculer grâce à une régularisation entropique comme une suite de problème de transport non-équilibrés.

  • Date:  november, 25th
  • Time: 14:00
  • Room: IRISA Rennes Salle Métivier (and visio)
  • Speaker:Kilian Fatras (PhD defense)
  • Title: OT and Deep Learning: Learning from one another
  • Abstract: Deep learning (DL) models are artificial neural networks and they have arisen as the current most competitive method to make data-driven decisions. In classification, these networks have a more complex representation of data and thus they make more complex predictions. However, DL’s recent successes are also due to the development of some mathematical fields : this thesis is about studying the different interactions of DL with one of these fields called Optimal Transport (OT).To measure the distance between probability distributions, one can rely on the OT theory. It defines a measure through the minimal displacement cost of a distribution to another. Its strength is to use the space geometry with a given ground cost on the data space. Several DL methods are built upon this theory.This thesis proposes to explore two faces of the interaction between OT and DL. We will first focus on how OT can define meaningful cost functions for neural networks. We will focus on how to define an OT based regularization for learning with noisy labels and how we can use OT to generate misclassified data for a pre-trained classifier. We will then explore what can be learnt about OT from DL applications, with a focus on the minibatch approximation of OT. We will answer what are the gains and downsides of the minibatch formulation both in theory and practice.
    Jury : Marco Cuturi / Christian WolfJury / Laetitia Chapel / Adam M. Oberman / Michèle Sebag

  • Date:  november, 18th
  • Time: 11h30
  • Room: D-001 (bat ENSIBS)
  • Speaker: Joachim Nyborg (PhD candidate, dpt Computer science, Aarhus Univ, Danemark)
  • Title: Unsupervised Cross-Region Adaptation by Temporal Shift Estimation
  • Abstract: The recent developments of deep learning models that capture the complex temporal patterns of crop phenology have greatly advanced crop classification of Satellite Image Time Series (SITS). However, when applied to target regions spatially different from the training region, these models perform poorly without any target labels due to the temporal shift of crop phenology between regions. To address this unsupervised cross-region adaptation setting, existing methods learn domain-invariant features without any target supervision, but not the temporal shift itself. As a consequence, these techniques provide only limited benefits for SITS. In this paper, we propose TimeMatch, a new unsupervised domain adaptation method for SITS that directly accounts for the temporal shift. TimeMatch consists of two components: 1) temporal shift estimation, which estimates the temporal shift of the unlabeled target region with a source-trained model, and 2) TimeMatch learning, which combines temporal shift estimation with semi-supervised learning to adapt a classifier to an unlabeled target region. We also introduce an open-access dataset for cross-region adaptation with SITS from four different regions in Europe. On this dataset, we demonstrate that TimeMatch outperforms all competing methods by 11% in F1-score across five different adaptation scenarios, setting a new state-of-the-art for cross-region adaptation.

  • Date:  october, 14th
  • Time: 11h30
  • Room: D-001 (bat ENSIBS)
  • Speaker :  Hugo Gangloff (Post doc OBELIX)
  • Title: Generalized pairwise Markov models and applications to unsupervised
    image segmentation
  • Abstract: Probabilistic graphical models such as hidden Markov models have found many applications in signal processing. In this presentation, we study on a particular extension of these models, the pairwise Markov  models. We follow an approach which enables us to combine pairwise Markov models with recent architectures from machine learning such as deep neural networks. We focus on the parameter estimation in these hybrid models by an approach based on the Expectation-Maximization algorithm linked with additional constraints for the interpretability of the hidden process to recover. As an illustration, the models will be applied to unsupervised image segmentation, especially when a correlated noise corrupts the image.

  • Date:  September, 24th (FRIDAY)
  • Room Ile Bailleron
  • Title: Seminaire au vert! (Annual team seminar)
  • Program

 

Comments are closed.