Seminars – 2020-2021

Previous seminars – 2020-2021

• Date: May, 21st (Friday)
  
• Time: 13:00
• Room : visio
• Speaker: Ahmed Nassar (PhD OBELIX & ETH Zurich) – PhD defense
  
• Title:  Learning to map street-side objects using multiple views
  
• Abstract: Creating inventories of street-side objects and their monitoring in cities is a labor-intensive and costly process. Field workers are known to conduct this process on-site to record properties about the object. These  properties can be the location, species, height, and health of a tree as an example. To monitor cities, gathering such information on a large scale becomes challenging. With the abundance of imagery, adequate coverage of a  city is achieved from different views provided by online mapping services (e.g., Google Maps and Street View, Mapillary). The availability of such imagery allows efficient creation and updating of inventories of street-side objects status by using computer vision methods such as object detection and multiple object tracking.
  
  This thesis aims at detecting and geo-localizing street-side objects, especially trees and street signs, from  multiple views. Solving the problem using an object detector, as with any problem solved with computer vision, brings up the usual problems of invariances such as occlusion, lighting, pose, viewpoint, and background. We rely on multiple views coupled with coarse pose information to solve these problems and for the benefit of getting more information about the object from these different views.

  Using multiple views brings another challenge, namely how to re-identify the objects in these different views to aggregate the information and not get duplicates of a particular object. Another major challenge is that the data sets acquired or used in our work contain imagery captured at a larger baseline, contrary to other data sets employed for person re-identification or self-driving and made of sequences of video frames. We propose several deep learning-based approaches to better detect, re-identify, and geo-localize objects and tackle these different challenges. In our first proposed approach, we aimed at investigating if using soft geometric constraints coupled with image evidence would provide a better re-identification or matching accuracy of objects across different views to overcome our large baseline obstacle. This method relied on image crops of the objects from ground-level imagery and geometric  metadata acquired from the image and then given as an input to a novel Siamese convolutional neural network-based architecture that matches the image crops. Having confirmed that infusing our model with soft  geometric constraints proved beneficial, our second approach aimed at achieving the same objective through an end-to-end model. The model takes as input a full image instead of crops, and our output is geo-localized bounding box detections tagged with identities across different views. To achieve such a task, we had to build a tool to annotate and create a data set of urban trees. Our final approach introduces another end-to-end model that relies on graph neural networks to improve flexibility and efficiency compared to the previous one. Also, in this approach, we include aerial imagery as another input for the first time.
  For all three proposed approaches in this thesis, we perform extensive experiments on curated data sets to demonstrate the proposed systems’ effectiveness.

  Keywords: Deep Learning, Computer Vision, Object detection, Re-identification, Graph Neural Networks, Urban objects, Multi-view

• Date: May, 6th
• Time: 11:30
• Room : visio
• Speaker:  Kilian Fatras (PhD OBELIX)
• Title:  Unbalanced minibatch Optimal Transport; applications to Domain Adaptation
  
• Abstract: Optimal transport distances have found many applications in machine learning  for their capacity to compare non-parametric probability distributions. Yet  their algorithmic complexity generally prevents their direct use on large scale  datasets. Among the possible strategies to alleviate this issue, practitioners  can rely on computing estimates of these distances over subsets of data, i.e. minibatches. While computationally  appealing, we highlight some limits of  this strategy, arguing it can lead to undesirable smoothing effects. As an  alternative,  we suggest that the same minibatch strategy coupled with unbalanced optimal transport  can yield more robust behavior. We discuss the associated theoretical properties,  such as unbiased estimators, existence  of (Clarke) gradients and concentration bounds.  Our experimental study shows that in challenging problems associated to  domain adaptation, the use of unbalanced optimal transport leads to significantly  better results, competing with or surpassing recent baselines.

• Date: April, 15th
• Time: 11:30
• Room : visio
• Speaker:  Clément Dechesne (post doc CNES, Toulouse)
• Title: Prédire et qualifier une segmentation sémantique avec les réseaux Bayésiens.
• Abstract: Bien que les méthodes de Deep Learning produisent les meilleurs résultats dans presque toutes les tâches de classification, elles produisent toujours des décisions trop confiantes et ne peuvent pas évaluer la pertinence de leurs prédictions. En effet, d’une part, il est facile de produire des images (non reconnaissables par les humains) que les réseaux existants jugent reconnaissables avec une grande confiance. D’un autre côté, un petit changement dans l’image d’entrée peut conduire à une prédiction très différente, toujours avec une confiance élevée. Aucune mesure de l’incertitude de la prédiction n’est fournie à partir des architectures de réseau actuelles. Il est possible de générer des estimations de probabilité pertinentes assimilées a une mesure de la confiance du modèle. Cependant, ces mesures sont basées sur des probabilités softmax qui ne permettent pas de saisir pleinement l’incertitude.
  L’apprentissage profond bayésien pour la segmentation sémantique permet de fournir une certaine mesure de l’incertitude dans la prédiction. Il peut être vu comme un ensemble de réseaux de neurones profonds, chacun fournissant une seule prédiction. Ainsi, il est possible de produire des cartes d’incertitude pertinentes permettant d’évaluer les performances du modèle en plus de la segmentation sémantique. Ces cartes d’incertitude permettent d’associer à un label un niveau de confiance, permettant par exemple de ne pas prédire les pixels incertains, ou encore la correction semi-automatique de base de données

• Date: April, 1st
• Time: 11:30
• Room : Visio
• Speaker:  Javiera Castillo (PhD OBELIX / ONERA)
• Title: Semi-supervised Semantic Segmentation in Earth Observation
  
• Abstract: The development of semi-supervised learning methods is essential to Earth Observation (EO) applications. Indeed, labeled remote sensing data are scarce and insufficient to train fully supervised models with good generalization capacities. Conversely, raw data are abundant and therefore it is crucial to leverage unlabeled inputs to build better models. This talk presents some approaches to perform semi-supervised semantic segmentation in Earth observation. First, we focus on the study of semi-supervised methods based on multi-task approaches, where an auxiliary task is created to leverage unlabeled data during the training of neural networks. We present our results in different benchmarks and compare the effect of different auxiliary tasks and loss functions to optimize. The second part concentrates on generative approaches for semi-supervised learning. More precisely, we investigate joint energy-based models, that are able to learn a discriminative task together with data distribution, and their extension to semi-supervision in Earth observation.

• Date: March, 18th
• Time: 11:30
• Room : Visio
• Speaker: Huy Tran (PhD OBELIX)
• Title: Continuous CO-Optimal Transport: A relaxation of Gromov-Wasserstein distance
  
• Abstract: In practical applications, the Gromov-Wasserstein (GW) distance provides a way to compare similarity matrices. However, this associates to two potential limitations: first, the similarity matrix is not always computationally feasible and necessarily meaningful, and second, to the limit, GW distance is only able to compare square matrices. CO-Optimal Transport (COOT) resolves these drawbacks and allows matrices of arbitrary size as inputs. On the other hand, its formulation remains only available in the discrete case. The very first continuous version of COOT is known as GW’s third lower bound. Despite of its practical usefulness, little has been known about its theoretical properties.In this presentation, we formulate the COOT problem in the general case (thus unifies both GW’s lower bound and COOT), under both balanced and unbalanced settings. Then we present some preliminary results, namely metric and convergence properties, and illustrate its application in heterogeneous domain adaptation. Finally, we discuss some open questions on various aspects: theoretical, statistical, numerical, as well as some potential applications.

• Date: December, 16 (WEDNESDAY) // joint seminar with Copernicus Master in Digital Earth
  
• Time: 11:30
• Room : visio
• Speaker: Ronan Fablet (LabSTICC/TOMS, IMT Atlantique, Brest)
• Title: Bridging physics and deep learning fo geophysical dynamics: application to ocean monitoring and surveillance
  
• Abstract: Whereas model-driven approaches represent the state-of-the-art for the analysis, simulation and reconstruction of physical systems, learning-based and data-driven frameworks become relevant schemes for a large number of application domains, including for the study of phenomena governed by physical laws. They offer new means to take advantage of the potential of observation and/or simulation big data. In this context, making the most of model-driven and data-driven paradigms naturally arises as a key challenge.
  In this talk, we will discuss these research avenues and give some illustrations on ocean monitoring applications (e.g., reconstruction of sea surface tracers, maritime traffic surveillance). We will specifically address how neural networks can provide novel means for the data-driven identification of representations of dynamical systems, which are imperfectly observed (e.g., noisy data, partial observation, irregular sampling..).
  Selected references: (full paper available on researchgate.net )
  R. Fablet et al. Joint learning of variational representations and solvers for inverse problems with partially-observed data. Arxiv, 2020.
  R. Lguensat et al. The Analog Data Assimilation. MWR 2017.
  V.D. Nguyen, R. Vadaine, G. Hajduch, R. Garello, R. Fablet. Multi-task Learning for Maritime Traffic Surveillance from AIS Data Streams. IEEE DSAA, 2018.V.D.
  Nguyen, S. Ouala, L. Drumetz, R. Fablet. EM-like Learning Chaotic Dynamics from Noisy and Partial Observations. arXiv preprint arXiv:1903.10335. 2019.
  S. Ouala, V.D. Nguyen, (…), R. Fablet. Learning Latent Dynamics for Partially Observed Systems. Chaos, 2020.
  F. Rousseau, L. Drumetz, R. Fablet. Residual Networks as Geodesic Flows of Diffeomorphisms. JMIV, 2019.

• Date: December, 16 (WEDNESDAY)
  
• Time: 14h00
• Room : Visio here
• Speaker: Caglayan Tuna (PhD student, OBELIX) – PhD Defense
• Title: Morphological Hierarchies for Satellite Image Time Series
  
• Abstract:  Although morphological hierarchies are today a well-established framework for single frame image processing, their extension to time-related data remains largely unexplored. This thesis aims to tackle the analysis of satellite image time series with tree-based representations. To do so, we distinguish between three kinds of models, namely spatial, temporal and spatial-temporal hierarchies. For each model, we propose a streaming algorithm to update the tree when new images are appended to the series. Besides, we analyze the structural properties of the different tree building strategies, thus requiring some projection methods for the spatio-temporal tree in order to obtain comparable structures. Then, trees are compared according to their node distribution, filtering capability and cost, leading to a superiority of the spatio-temporal tree (a.k.a. space-time tree). Hence, we review spatio-temporal attributes, including some new ones, that can been extracted from the space-time tree in order to compute some multiscale features at the pixel or image level. These attributes are finally involved in tools such as filtering and pattern spectrum for various remote sensing based applications.

• Date: December, 4
  
• Time: 9h30
• Room : RENNES 2 Visio
• Speaker: Romain Tavenard (LETG Rennes & OBELIX, IRISA) – HDR Defense
• Title:  Apprentissage statistique et séries temporelles
  
• Abstract: A Jupyter Book version of the thesis is available online : https://rtavenar.github.io/hdr

• Date: December, 3 (THURSDAY) // joint seminar with Copernicus Master in Digital Earth
  
• Time: 11:30
• Room : VISIO
• Speaker: Damien Arvor (CR CNRS, LETG, Rennes2)
• Title: Monitoring agricultural dynamics in the southern brazilian Amazon with remote sensing data
  
• Abstract: Les dynamiques agricoles dans le sud de l’Amazonie brésilienne impactent fortement les écosystèmes naturels. Bien que la déforestation ait largement baissé depuis 2004, les processus d’expansion, d’intensification et de diversification agricole se poursuivent et soulèvent de nouvelles interrogations, notamment dans un contexte de changement climatique. L’imagerie satellitaire et ses produits dérivés (e.g. cartes d’occupation du sol) apportent des informations précieuses pour apporter des éléments de réponse à ces questions de recherche. Par exemple, les séries temporelles d’indice de végétation sont utilisées pour cartographier les pratiques agricoles alors que les estimations de précipitations permettent de spatialiser les premières évidences du changement climatique. Toutefois, les méthodes de traitement d’images satellitaires favorisées aujourd’hui (e.g. machine learning) présentent certaines limitations qui peuvent gêner l’interprétation de l’information spatiale par des utilisateurs aux objectifs divers et variés (géographes, écologues, climatologues, etc). Il convient donc de discuter de l’intérêt de méthodes « kwowledge-driven » pour l’interprétation des images satellitaires, en soulignant notamment l’apport des ontologies.

• Date: November, 5
  
• Time: 9h30
• Room : visio (link : if you missed the defense and you want to see it )
• Speaker: Titouan Vayer (OBELIX, IRISA Vannes) – PhD Defense
• Title: A contribution to Optimal Transport on incomparable spaces
  
• Abstract: Optimal Transport is a theory that allows to define geometrical notions of distance between probability distributions and to find correspondences, relationships, between sets of points. Many machine learning applications are derived from this theory, at the frontier between mathematics and optimization. This thesis proposes to study the complex scenario in which the different data belong to \emph{incomparable} spaces. In particular we address the following questions: how to define and apply the optimal transport between graphs, between structured data? How can it be adapted when the data are varied and not embedded in the same metric space? This thesis proposes a set of Optimal Transport tools for these different cases. An important part is notably devoted to the study of the Gromov-Wasserstein distance whose properties allow to define interesting transport problems on incomparable spaces. More broadly, we analyze the mathematical properties of the various proposed tools, we establish algorithmic solutions to compute them and we study their applicability in numerous machine learning \emph{scenarii} which cover, in particular, classification, simplification, partitioning of structured data, as well as heterogeneous domain adaptation.

• Date: October, 23 (FRIDAY)
  
• Time: 9h30-17h00
• Room :  A105-107, ENSIBS, Vannes
• Title: Seminaire au vert! (Annual team seminar)
• Program here

• Date: October, 1st
  

• Time: 11:30
• Room : amphi A 101 / DSEG
• Speaker: Badie Belmouhcine (post doc IFREMER / IRISA-OBELIX)
• Title: Generative Adversarial Networks for 3D voxelgrid generation
• Abstract: Generative Adversarial Networks are successful and powerful generative models, that were applied successfully for many tasks such as image and video generation, style transfer, in-painting, domain adaptation, etc. The generation of 3D shapes is particularly challenging, due to the high dimensional input space and the lack of training data.
  During my internship in the CNAM, we studied the use of Generative Adversarial Networks for 3D voxelized object generation and how we can improve their performance by leveraging some properties of 3D voxelized shapes. We first showed that the training of Wasserstein GANs for 3D shape generation can be significantly improved by combining the Gradient Penalty with Spectral Normalization. We also proposed to adapt the model to match the requirement of binary voxel generation, by employing softargmax in the generator. Then, we modified the cost function to enforce the plane symmetry of the generated objects. Finally, since 3D voxelized objects are more characterized by the structure than by the texture, we started to evaluate the use of Self-Attention GANs and Non-local networks in the 3D context, by modeling long range dependencies across different regions of the voxel grid.

• Date: Septembre, 23rd (WEDNESDAY)
  
• Time: 11:00
• Room : Salle du conseil / bât. Présidence – campus de Tohannic
• Speaker:  Lucas Drumetz (MCF Traitement du signal, IMT Brest)
• Title: Learning Dynamical Systems from data, application to multi and hyperspectral time series
• Abstract: The evergrowing amount of remote sensing data acquired from satellites make it possible to access historical catalogs of image data for multiple applications, e.g. for ocean altimetry or terrestrial optical images. If data assimilation techniques have been extensively used to take advantage of those observations for the former, the time dimension, however, is often neglected in some data processing pipelines for the latter (e.g. classification and unmixing), even when the end-goal is the characterization of the dynamics of the imaged scene. This is mainly due to the fact that the data is partial and noisy. In this talk, I will present a few machine learning techniques and methods allowing to learn dynamical systems from data, mainly using dedicated neural network architectures. I will present some associated methodological challenges (dealing with chaotic systems, stochastic systems, partial data…) and show a few illustrations on toy datasets and a detailed application to simulated and real time-series of satellite multi/hyperspectral data.