Seminars – 2019

The seminars of OBELIX team are currently held on Mondays 11:00 am, every two weeks, at the IRISA lab, Tohannic campus, room D106.

The seminar is coordinated by Chloé Friguet.

Previous seminars – 2019

  • Date: July, 11 & 12 : Seminaire au vert! (Annual team seminar)
  • Room: Quiberon 🙂
  • program here

  • Date: June, 17
  • Room : D106
  • Speaker: ValĂ©rie Garès (MCF Statistique, INSA / IRMAR, Rennes)
  • Title: On the use of optimal transportation theory to recode variables and application to database merging
  • Abstract: When databases are constructed from heterogeneous sources, it is not unusual that different encodings are used for the same outcome. This work considers the problem of finding a relevant way to recode a categorical variable before merging two databases. The method is an application of optimal transportation  where we search for a bijective mapping between the distributions of such variable in two databases. Using that common covariates appear in the two databases, the objective is to minimize the expectation of a cost function reflecting a distance measure in the space of the covariates. The first form of the algorithm needs the assumption that the covariates may follow the same distribution in the two databases [1]. We proposed different models stating a novel approach to answer the problem and relaxing this hypothesis. Our different models are compared in a simulation study in different scenarios and are applied to a real dataset.
    Ref: Dimeglio C*, Garès V.*, Kosorok M. R., Guernec G., Fantin R., Lepage B. and Savy N. On the use of optimal transportation theory to merge databases. Application to clinical trials. En révision.

  • Date: May, 13
  • Room : salle des conseils /Bat. de la PrĂ©sidence
  • Speaker: ClĂ©ment Dechesne (Post-doc, IMT Atlantique / Lab-STICC, Brest)
  • Title: Ship detection and characterization in Sentinel-1 SAR images with multi-task deep learning
  • Abstract: The detection of inshore and offshore ships is an important issue in both military and civilian fields. It helps monitoring fisheries, managing maritime traffics, ensuring safety of coast and sea, etc. In operational contexts, ship detection is traditionally performed by a human observer who identifies all kinds of ships from visual analysis on remotely-sensed images. Such a task is very time consuming and cannot be conducted at a very large scale, while Sentinel-1 SAR data now provides regular, worldwide coverage. Meanwhile, with the emergence of GPUs, deep learning methods are now established as state-of-the-art solutions for computer vision, replacing human intervention in many contexts. They have been shown to be adapted for ship detection, most often with very high resolution SAR or optical imagery. We go one step further and propose a deep neural network for the joint detection, classification and length estimation of ships from SAR Sentinel-1 data. We benefit from synergies between AIS (Automatic Identification System) and Sentinel-1 data to build significant training datasets. We then design a multi-task neural network architecture composed of one joint convolutional network connected to three task-specific networks, namely for ship detection, classification and length estimation. The experimental assessment showed our network provides satisfactory results, with accurate classification and length estimation.

  • Date: April, 29
  • Room : D106
  • Speaker: Ahmed Nassar (Ph.D student, OBELIX-IRISA, Vannes)
  • Title: RegisTree |  Simultaneous multi-view instance detection with learned geometric soft-constraints
  • Abstract: RegisTree is an automated system that can detect and inventory street trees automatically. From publicly available imagery (Google maps and Google Street View at the moment) of a city, our system produces a list of the location, species and trunk diameter of each street tree. Our approach uses deep learning to identify the location of a tree, classify its species, and approximate its trunk diameter. It combines publicly available geo-referenced Google Maps aerial and street view images along with map data to provide a comprehensive and accurate catalogue of street trees. We have created an annotation tool, a multi-view ReID dataaset, and a multi-view object detection method.

  • Date: April, 05 (14h00 – extra seminar)
  • Room : D106
  • Speaker: Xiaofang Wang (R&D Engineer, INRIA-Rainbow, Rennes)
  • Title:Weakly Supervised Learning for Object Detection
  • Abstract: Machine learning and Deep convolutional networks have become the standard pipeline for a variety of computer vision task such as image classification, object detection, segmentation, etc. These models are known to be data-driven, they requires huge-volume of data labelled by human being or experts. However, manual labeling task is very tedious, time-consuming and expensive, learning from “weaker” annotations (e.g. only image-level category labels to localize object instances by a bounding box) become very interesting to extend the fully-supervised learning algorithms to new applications. In this seminar, I will present two of my methods of weakly supervised learning methods for object detection, including its motivation, originality, evaluation metrics, and future ideas. The first one is to enhancing DPM for weakly supervised learning and The second one is weakly supervised deep learning for crowd detection and counting.

  • Date: April, 01
  • Room : D106
  • Speaker: Romaric Gaudel (MCF ENSAI CREST / LACODAM, IRISA, Rennes)
  • Title: Recommendation as a sequential process
  • Abstract: The main track on Recommender Systems (RS) looks for the best next recommendation. At the  opposite, few papers take into account the fact  that a RS handles sequences of recommendations: it  recommends some items  to a user, it gathers the corresponding feedback, it recommends some  other items (to the same or another user), it gathers corresponding  feedback, etc. In my presentation I review a set of  approaches handling two consequences of that aspect of RS: (i) you can reduce computation time by relying on that sequence, (ii) you have to explore to get the best from your RS.

  • Date: March, 20 (14h00 – extra seminar)
  • Room : D106
  • Speaker: Geoffrey Roman-Jimenez (Post-doctoral researcher, Institut de Recherche en Informatique, Toulouse)
  • Title: Science des donnĂ©es et apprentissage statistique pour l’extraction de connaissances Ă  partir de donnĂ©es et d’images
  • Abstract: Dans le cadre de ce sĂ©minaire, je prĂ©senterai une partie de mes travaux de recherche portant sur l’utilisation de mĂ©thodes de sciences des donnĂ©es et d’apprentissage statistique (machine/deep-learning, modĂ©lisation/fouille/visualisation de donnĂ©es) pour l’extraction de connaissances Ă  partir de donnĂ©es et d’images. Les premiers travaux que je prĂ©senterai porteront sur l’utilisation du transfert d’apprentissage pour la dĂ©tection de structure au sein de documents manuscrits historiques, en l’absence de vĂ©ritĂ©-terrain. ’exposerai ensuite mes travaux sur la rĂ©identification de vĂ©hicules par l’utilisation des espaces latents de rĂ©seaux convolutifs profonds. Je finirai sur la prĂ©sentation des diffĂ©rents travaux que je mène actuellement sur la modĂ©lisation, la fouille et la visualisation de (mĂ©ta-)donnĂ©es hĂ©tĂ©rogènes issues des systèmes de vidĂ©oprotection.

  • Date: March, 18 (11h00 – usual seminar)
  • Room : D106
  • Speaker: Yichang Wang (Ph.D. student OBELIX / LACODAM, IRISA, Rennes)
  • Title: Learning Interpretable Shapelets for Time Series Classification through Adversarial Regularization
  • Abstract: Times series classification can be successfully tackled by jointly learning a shapelet-based representation of the series in the dataset and classifying the series according to this representation. However, although the learned shapelets are discriminative, they are not always similar to pieces of a real series in the dataset. This makes it difficult to interpret the decision, i.e. difficult to analyze if there are particular behaviors in a series that triggered the decision. In this paper, we make use of a simple convolutional network to tackle the time series classification task and we introduce an adversarial regularization to constrain the model to learn more interpretable shapelets. Our classification results on all the usual time series benchmarks are comparable with the results obtained by similar state-of-the-art algorithms but our adversarially regularized method learns shapelets that are, by design, interpretable.

  • Date: March, 18 (13h00 – extra seminar)
  • Room : A106
  • Speaker: Lina Fahed (Post-doctoral researcher, IMT-Atlantique, Lab-STICC/DECIDE, Brest)
  • Title: Prediction and influence of the occurrence of events in complex sequences
  • Abstract: For several years now, a new phenomenon related to digital data is emerging: data which is increasingly voluminous, varied and rapid, appears and becomes available, it is often referred to as complex data. Complex data is very rich in hidden information, which increases its predictive and detection power. For many years, prediction and detection models in the data mining literature have been proposed for so-called simple data (from a single source, with a clear structure, etc.). However, the generalization and predominance of complex data lead to new challenges, mainly related to taking into account the new characteristics of this data in the modeling process. Consequently, proposing new models adapted to complex data has become essential to better detect, predict and even influence the future, which represents the main challenge that I have chosen to address during my research. This presentation covers my contributions for mining complex data, with a focus on complex sequences, by proposing models for: (i) events prediction, (ii) emergence detection, (iii) events influence and (iv) anomaly detection.

  • Date: March, 11
  • Room : D106
  • Speaker: François Septier (Pr. Statistique / Univ. Bretagne-Sud – Lab-STICC, Vannes)
  • Title: Advanced Monte-Carlo methods for Bayesian filtering
  • Abstract: After a brief description of the general principle of Monte-Carlo methods for Bayesian filtering, I will present some recent contributions in this field. More specifically, I will firstly describe a novel estimator affiliated with the normalized importance sampling estimator but using two proposition densities rather than a single one. I will then present a revisit of the importance sampling with resampling mechanism as a whole in order to produce Monte Carlo samples that are independent, contrary to the classical mechanism, which enables us to develop two new estimators. Finally, I will discuss some sequential adaptations of the Markov Chain Monte Carlo algorithm which represent an interesting alternative to these importance sampling based techniques.

  • Date: February, 25 March, 4
  • Room : D106
  • Speaker: Adan Salazar (Geo-intelligence remote sensing group, Mexico)
  • Title: Challenges and opportunities in relation to Sargassum blooms along the coast of the Mexican Caribbean
  • Abstract: Macroalgal blooms are becoming more frequent events worldwide in response to several factors, some of which have not been proven as responsible for these common events. Reports in the scientific literature for the Atlantic and Pacific coast for Golden tides or  looms of the genus Sargassum have appeared recently. Past and  recent (2011, 2015, 2018) blooms formed by Sargassum species along the Mexican Caribbean coast make us aware  of the necessity to tackle this events not only locally, but at a Regional and Global scale. We have identified at least three species of Sargassum comprising the blooms occurring along the coast of the Mexican Caribbean.  Monthly collections of this biomass derived in taxonomic information, in which other benthic species were also  identified. The chemical characterization of the collected biomass will provide basic information for Sargassum  valorization. Moreover, based on satellite image analysis, we followed the event in order to quantify its pelagic volume and the possible impact in touristic areas and vulnerable ecosystems in nearby coastal area of Quintana  Roo. Basic and applied information on Sargassum biomass may contribute to the understanding of these events  and also suggest management and possible solutions to their widespread occurrence worldwide.

  • Date: February, 11
  • Room : D106
  • Speaker: Marc Russwurm (Ph.D. student / Univ. Munich, Germany)
  • Title: End-to-end Learning for Early Classification of Time Series
  • Abstract: Classification of time series is a topical issue in machine learning. While accuracy stands for the most important evaluation criterion, some applications require decisions to be made as early as possible. Optimization should then target a compromise between earliness, i.e., a capacity of providing a decision early in the sequence, and accuracy. During my stay at IRISA-Obelix, I helped designing a generic, end-to-end trainable framework for early classification of time series. This framework embeds a learnable decision mechanism that can be plugged into a wide range of already existing models. We present results obtained with deep neural networks on a diverse set of time series classification problems. The developed approach compares well to state-of-the-art competitors while being easily adaptable by any existing neural network topology that evaluates a hidden state at each time step.

  • Date: January, 28
  • Room : C011
  • Speaker: Behzad Mirmahboub (post-doc / OBELIX-IRISA, Vannes)
  • Title: Feature Design and Metric Fusion in Person Re-Identification
  • Abstract: Person re-identification is the problem of recognizing a person between several non overlapped cameras. It has important applications in surveillance systems and can reduce human labor and errors of matching persons. Nevertheless, matching two images of a person with different views, poses and illuminations in presence of occlusion and noise is very challenging. Various types of features and metrics are proposed in order to image retrieval for person re-id. In our proposed method we extract different types of features from images and obtain several ranking lists based on their distances. Then, we combine those lists according to their confidence to find the best ranking list.
    Reference: Person re-identification by order-induced metric fusion (Neurocomputing 2018)


Comments are closed.