HAL-Ecole des Ponts ParisTech
Not a member yet
42743 research outputs found
Sort by
Global transport and evolution of the Hunga sulfates and water from satellite and ground-based observations: a 2-year aftermath
No full textInternational audienceThe eruption of the submarine Hunga volcano on 15 January 2022 was associated with a powerful blast that injected water and sulfur throughout the stratospheric layer, leading to an unprecedented increase in the global stratospheric water mass by 13% as compared to climatological levels and a 5-fold increase in the stratospheric aerosol load, the highest in the last three decades. The unique nature and magnitude of the global stratospheric perturbation by the Hunga eruption ranks it among the most remarkable climatic events in the modern observation era.In this study, we combine the data from various satellite instruments (MLS, OMPS-LP, CALIOP, SAGE III, ACE-FTS) together with ground-based lidars and balloon soundings at various locations to investigate the global transport and evolution of the Hunga sulfates and moisture during more than two years following the event. In particular, we address the following aspects: timescales of meridional and vertical progression of the moisture and aerosol-rich layers; bulk transport into the southern high-latitudes; episodes of in-mixing into the northern extratropics; growth and sedimentation of Hunga sulfates. In addition, we compare the Hunga stratospheric aerosol perturbation with those produced by moderate and major eruptions in terms of the e-folding time, spatial evolution and aerosol microphysical parameters
The CATMuS initiative: building large and diverse corpora for handwritten text recognition
No full textInternational audienc
Straightforward Way to Provide Antibacterial Properties to Healthcare Textiles through N -2-Hydroxypropyltrimethylammonium Chloride Chitosan Crosslinking
No full textInternational audienceA straightforward and cost-effective way to coat polypropylene fibers, designed for healthcare textiles, was developed through chitosan crosslinking for antibacterial purposes. As polypropylene is an inert material, the goal was to physically trap the inert fibers through a network of crosslinked chitosan or a quaternized derivative (to enhance the antibacterial action). First, chitosan or its quaternized derivative was physically deposited by impregnation or spraying. Then, chitosan was crosslinked in glutaraldehyde solution followed by its quaternized derivative with a diisocyanate. Coated fabrics were characterized by infrared spectroscopy (IR), weight gain measurements, and scanning electron microscopy (SEM) coupled with energy dispersive X-ray (EDX). This led us to conclude that spraying provides uniform deposition while maintaining the fabric’s porosity. Acidic washing allowed us to prove that chitosan and its quaternized derivative were successfully immobilized on the fabric. Biological assays were conducted against two major strains of bacteria responsible for nosocomial infections: Staphylococcus aureus (Gram-positive) and Escherichia coli (Gram-negative). Chitosan-crosslinked samples did not show significant antibacterial activity, but the quaternized derivative allowed a significant decrease in S. aureus colonization. This study presents a simple and scalable process to coat inert fabrics with a polymer containing reactive functions potentially available to graft various additional antimicrobial agents
Calcul efficace des coefficients de transport en dynamique moléculaire
No full textIn this thesis we present some new methods for efficiently computing transport coefficients (TCs) with molecular dynamics. Most standard numerical methods currently used to approximate TCs are plagued by large errors, particularly statistical errors. Although there have been attempts to devise more efficient numerical methods and reduce the associated errors, most people still resort to direct, brute-forced methods which are computationally cumbersome and largely inefficient. The techniques we present focus on reducing the variance associated with such computations, and are presented in three chapters. In chapter one, we propose the use of synthetic forcings to increase the linear response regime in nonequilibrium molecular dynamics simulations, which acts as a proxy for reducing the variance associated with the estimator at hand. Chapter two introduces a method based on transient dynamics, employing the use of couplings as a control variate based on the so-called subtraction technique by Ciccotti and Jacucci (1975). Finally, chapter three explores equilibrium techniques based on fluctuation formulas (such as the Green--Kubo and Einstein formulas), providing extensive numerical analysis of various improved estimators and proposing neural networks (NNs) as a variance reduction tool. Here, NNs are used to approximate the solution of a Poisson equation, which serves as a control variate in the TC computation at hand.Dans cette thèse, nous présentons de nouvelles méthodes pour calculer efficacement les coefficients de transport (CT) avec la dynamique moléculaire. La plupart des méthodes numériques standards actuellement utilisées pour approximer les CT sont entachées d'erreurs importantes, en particulier d'erreurs statistiques. Malgré des tentatives pour concevoir des méthodes numériques plus efficaces et réduire les erreurs associées, la plupart des chercheurs ont encore recours à des méthodes directes, qui sont lourdes en termes de calcul et largement inefficaces. Les techniques que nous présentons sont axées sur la réduction de la variance associée à ces calculs et sont présentées en trois chapitres. Dans le premier chapitre, nous proposons d'utiliser des forçages synthétiques pour augmenter le régime de réponse linéaire dans les simulations de dynamique moléculaire à l'équilibre, ce qui permet de réduire la variance associée à l'estimateur utilisé. Le deuxième chapitre présente une méthode basée sur la dynamique transitoire, qui utilise les couplages comme variable de contrôle selon la technique dite de soustraction de Ciccotti et Jacucci (1975). Enfin, le troisième chapitre explore les techniques d'équilibre basées sur les formules de fluctuation (telles que les formules de Green--Kubo et d'Einstein), en fournissant une analyse numérique approfondie de divers estimateurs améliorés et en proposant les réseaux neuronaux (NN) comme outil de réduction de la variance. Ici, les réseaux neuronaux sont utilisés pour approximer la solution d'une équation de Poisson, qui sert de variable de contrôle dans le calcul de CT en question
OccFeat: Self-supervised Occupancy Feature Prediction for Pretraining BEV Segmentation Networks
No full textInternational audienceWe introduce a self-supervised pretraining method, called OccFeat, for camera-only Bird's-Eye-View (BEV) segmentation networks. With OccFeat, we pretrain a BEV network via occupancy prediction and feature distillation tasks. Occupancy prediction provides a 3D geometric understanding of the scene to the model. However, the geometry learned is class-agnostic. Hence, we add semantic information to the model in the 3D space through distillation from a self-supervised pretrained image foundation model. Models pretrained with our method exhibit improved BEV semantic segmentation performance, particularly in lowdata scenarios. Moreover, empirical results affirm the efficacy of integrating feature distillation with 3D occupancy prediction in our pretraining approach
Estimation de la pose de nouveaux objets rigides
No full textObject pose estimation is an important computer vision problem as it has great impact in many applications such as robotics, augmented reality, and autonomous driving. While existing supervised object pose estimation methods have achieved remarkable performance, they heavily rely on extensive training data specific to each target object, and cannot generalize to novel objects. In this thesis, we address this problem of generalization, and propose scalable methods for detecting, segmenting, and estimating the 6D pose of novel objects in RGB images. Our methods require only the 3D model or a single reference image of test objects for inference, making them suitable for practical scenarios in real-world applications.Our methods are grounded on the concept of ``templates'', which are 2D views of target 3D objects. The idea of using templates for object pose estimation is not new: the first work on this seems to be Murase's and Nayar's in 1995. We nevertheless show that, thanks to recent advances in machine learning, in particular on unsupervised image features, templates can have very interesting properties to generalize to unseen objects efficiently: They do not require training while adapting to new objects; they can be matched extremely fast; they can be robust to occlusions thanks to high dimensional local image features.Our first method is CNOS, a simple yet powerful method for segmenting novel objects in RGB images from their 3D models. CNOS generates segmentation proposals from the input image using Segment Anything and matching them with templates rendered from the 3D models and represented with the foundation features of DINOv2. CNOS significantly outperforms existing methods on multiple datasets and has been awarded as the best method for 2D detection/segmentation of unseen object in the BOP challenge 2023. CNOS is also used as the default detection method for both model-based, model-free unseen object pose estimation at the BOP challenge 2024.We then reconsider in details existing template matching methods for pose estimation and analyse their limitations when dealing with novel objects. We introduce a new approach that matches the input testing image with templates from their 3D models. Our approach is trained on a limited number of reference objects with contrast learning and generalizes well to novel objects.We further extend this work in a method we call GigaPose, a novel approach that is significantly faster and more robust against input segmentation errors. GigaPose samples templates only on two degrees of freedom~(DoFs) for estimating out-of-plane rotation, then uses feature matching to estimate the remaining four DoFs. GigaPose seamlessly integrates with existing refinement methods. It achieves state-of-the-art results for RGB-based methods on the standard BOP benchmark while is 35 times faster comparing to existing methods for the coarse pose estimation stage.Finally, we present NOPE, a simple method to estimate novel object pose from a single reference image. Our approach takes a single image of a new object as input and predicts the relative pose of this object in new images without prior knowledge of the object’s 3D model. We achieve this by training a model to directly predict discriminative embeddings that resemble templates for viewpoints surrounding the object. This prediction is done using a simple U-Net architecture with attention and conditioned on the desired pose, which yields extremely fast inference. We compare our approach to state-of-the-art methods and show it outperforms them both in terms of accuracy and robustness on both synthetic and real-world datasets.L'estimation de pose d'objets est un problème important en vision par ordinateur, car elle a un grand impact sur de nombreuses applications telles que la robotique, la réalité augmentée et la conduite autonome. Bien que les méthodes existantes d'estimation de pose d'objets supervisées aient atteint des performances remarquables, elles dépendent beaucoup de données d'entraînement spécifiques à chaque objet et ne peuvent pas généraliser à de nouveaux objets qui n'ont pas été vus pendant l'entraînement. Dans cette thèse, nous abordons ce problème de généralisation et proposons des nouvelles méthodes pour détecter, segmenter et estimer la pose 6D pour des nouveaux objets dans des images RGB. Nos méthodes n'utilisent que le modèle 3D ou une seule image de référence des objets de test pour l'inférence, permettant des scénarios pratiques dans des applications réelles. Nos méthodes sont basées sur le concept de templates, qui sont des vues 2D d'objets 3D. L'idée d'utiliser des templates pour l'estimation de la pose d'objets n'est pas nouvelle: le premier travail à ce sujet est celui de Murase et Nayar en 1995. Cependant, nous montrons que, grâce aux récents progrès en apprentissage automatique, en particulier sur les caractéristiques d'images non supervisées, les templates peuvent avoir des propriétés très intéressantes pour se généraliser efficacement à des objets jamais vus auparavant: ils ne nécessitent pas d'entraînement pour s'adapter à de nouveaux objets; ils peuvent être appariés extrêmement rapidement; ils peuvent être robustes aux occultations grâce à des caractéristiques locales d'image. Les principales contributions de cette thèse sont les suivantes. Premièrement, nous introduisons CNOS, une méthode simple mais efficace pour segmenter de nouveaux objets dans des images RGB à partir de leurs modèles 3D. Notre méthode génère des hypothèses de segmentation à partir de l’image d’entrée en utilisant Segment Anything et les compare avec des templates, ici des images rendues des modèles 3D en utilisant les caractéristiques DINOv2. CNOS surpasse de manière significative les méthodes existantes sur plusieurs jeux de données dans le BOP Challenge 2023.Deuxièmement, nous revisitons les méthodes de correspondance de templates existantes et démontrons leurs limites lorsqu'il s'agit de nouveaux objets. Nous introduisons ensuite une nouvelle approche qui associe l'image de test d'entrée avec des templates de leurs modèles 3D. Notre approche est entraînée sur un petit nombre d'objet de référence avec un apprentissage par contraste et se généralise bien à de nouveaux objets. Troisièmement, nous introduisons GigaPose, une nouvelle approche qui est significativement plus rapide et plus robuste contre les erreurs de segmentation en entrée. GigaPose échantillonne des templates uniquement sur deux degrés de liberté pour estimer la rotation hors-plan, puis utilise la correspondance de caractéristiques pour estimer les quatre DoFs restants. GigaPose peut s'intégrer parfaitement avec les méthodes de raffinement existantes et atteindre des résultats à la pointe de la technologie sur les jeux de données standards BOP. Quatrièmement, nous présentons NOPE, une méthode simple pour estimer la pose d'objet nouveaux à partir d'une seule image de référence. Notre méthode prend une seule image d'un nouvel objet comme entrée et prédit la pose relative de cet objet dans de nouvelles images sans connaissance préalable du modèle 3D de l'objet. Nous y parvenons en entraînant un modèle à prédire directement des embeddings discriminatifs pour des points de vue entourant l'objet. Cette prédiction est effectuée en utilisant une architecture U-Net simple avec attention et conditionnée par la pose souhaitée, ce qui permet une inférence extrêmement rapide. Nous comparons notre approche aux méthodes éxistant et montrons qu'elle les surpasse en termes de précision et de robustesse sur des jeux de données synthétiques et réels
Prospective Surfaces et Interfaces Continentales : 2024-2028
No full textVoir aussi : Le colloque de prospective SIC s’est déroulé sur quatre demi-journées, du 05/02/2024 au 07/02/2024, à Saint-Malo : https://prospective-sic.sciencesconf.org/resource/page/id/1
The Many Channels of Firm's Adjustment to Energy Shocks: Evidence from France
No full textInternational audienceBased on firm-level data in the French manufacturing sector, we find that firms adapt quickly, strongly and through multiple channels to energy shocks, even though electricity and gas bills represent a small share of their total costs. Over the period 1996–2019, faced with an idiosyncratic energy price increase, firms reduce their energy demand, improve their energy efficiency, increase intermediate inputs imports and optimize energy use across plants. Firms are also able to pass-through the cost shock fully into their export prices. Their production, exports and employment fall. A consequence of these multiple adjustment mechanisms is that the fall in profits is either non-significant, small or specific to only the most energy-intensive firms. We also find that the impact of electricity shocks has weakened over time, suggesting that only firms able to adapt their production process to energy cost shocks have survived. Importantly, when faced with large electricity and gas price increases, firms are less able to reduce their consumption. These results shed light on the mechanisms of resilience of the European manufacturing sector in the context of the present energy crisis
Tropical reproducing kernels and optimization
No full textInternational audienceHilbertian kernel methods and their positive semidefinite kernels have been extensively used in various fields of applied mathematics and machine learning, owing to their several equivalent characterizations. We here unveil an analogy with concepts from tropical geometry, proving that tropical positive semidefinite kernels are also endowed with equivalent viewpoints, stemming from Fenchel-Moreau conjugations. This tropical analogue of Aronszajn's theorem shows that these kernels correspond to a feature map, define monotonous operators, and generate max-plus function spaces endowed with a reproducing property. They furthermore include all the Hilbertian kernels classically studied as well as Monge arrays. However, two relevant notions of tropical reproducing kernels must be distinguished, based either on linear or sesquilinear interpretations. The sesquilinear interpretation is the most expressive one, since reproducing spaces then encompass classical max-plus spaces, such as those of (semi)convex functions. In contrast, in the linear interpretation, the reproducing kernels are characterized by a restrictive condition, von Neumann regularity. Finally, we provide a tropical analogue of the "representer theorems", showing that a class of infinite dimensional regression and interpolation problems admit solutions lying in finite dimensional spaces. We illustrate this theorem by an application to optimal control, in which tropical kernels allow one to represent the value function