École nationale des ponts et chaussées

We present a new integral equation formulation of the polarizable continuum model (PCM) which allows one to treat in a single approach dielectrics of different nature: standard isotropic liquids, intrinsically anisotropic medialike liquid crystals and solid matrices, or ionic solutions. The present work shows that integral equation methods may be used with success also for the latter cases, which are usually studied with three-dimensional methods, by far less competitive in terms of computational effort. We present the theoretical bases which underlie the method and some numerical tests which show both a complete equivalence with standard PCM versions for isotropic solvents, and a good efficiency for calculations with anisotropic dielectrics.

We present the full implementation of the integral equation formalism (IEF) we have recently formulated to treat solvent effects. The method exploits a single common approach for dielectrics of very different nature: standard isotropic liquids, intrinsically anisotropic media like liquid crystals, and ionic solutions. We report here an analysis of its both formal and technical details as well as some numerical applications addressed to state the achieved generalization to all kinds of molecular solutes and to show the equally reliable performances in treating such different environmental systems. In particular, we report, for isotropic liquids, data of solvation free energies and static (hyper)polarizabilities of various molecular solutes in water, for anisotropic dielectrics, a study of an SN2 reaction, and finally, for ionic solution, a study of some structural aspects of ion pairing.

Monocular depth estimation, which plays a crucial role in understanding 3D scene geometry, is an ill-posed problem. Recent methods have gained significant improvement by exploring image-level information and hierarchical features from deep convolutional neural networks (DCNNs). These methods model depth estimation as a regression problem and train the regression networks by minimizing mean squared error, which suffers from slow convergence and unsatisfactory local solutions. Besides, existing depth estimation networks employ repeated spatial pooling operations, resulting in undesirable low-resolution feature maps. To obtain high-resolution depth maps, skip-connections or multilayer deconvolution networks are required, which complicates network training and consumes much more computations. To eliminate or at least largely reduce these problems, we introduce a spacing-increasing discretization (SID) strategy to discretize depth and recast depth network learning as an ordinal regression problem. By training the network using an ordinary regression loss, our method achieves much higher accuracy and faster convergence in synch. Furthermore, we adopt a multi-scale network structure which avoids unnecessary spatial pooling and captures multi-scale information in parallel. The proposed deep ordinal regression network (DORN) achieves state-of-the-art results on three challenging benchmarks, i.e., KITTI [16], Make3D [49], and NYU Depth v2 [41], and outperforms existing methods by a large margin.

We analyze a model of imperfect price competition between intermediation service providers. We insist on features that are relevant for informational intermediation via the Internet: the presence of indirect network externalities, the possibility of using the nonexclusive services of several intermediaries, and the widespread practice of price discrimination based on users' identity and on usage. Efficient market structures emerge in equilibrium, as well as some specific form of inefficient structures. Intermediaries have incentives to propose non-exclusive services, as this moderates competition and allows them to exert market power. We analyze in detail the pricing and business strategies followed by intermediation services providers. Copyright 2003 by the RAND Corporation.

This paper presents a method to automatically and efficiently detect face tampering in videos, and particularly focuses on two recent techniques used to generate hyper-realistic forged videos: Deepfake and Face2Face. Traditional image forensics techniques are usually not well suited to videos due to the compression that strongly degrades the data. Thus, this paper follows a deep learning approach and presents two networks, both with a low number of layers to focus on the mesoscopic properties of images. We evaluate those fast networks on both an existing dataset and a dataset we have constituted from online videos. The tests demonstrate a very successful detection rate with more than 98% for Deepfake and 95% for Face2Face.

Over the last years, deep convolutional neural networks (ConvNets) have transformed the field of computer vision thanks to their unparalleled capacity to learn high level semantic image features. However, in order to successfully learn those features, they usually require massive amounts of manually labeled data, which is both expensive and impractical to scale. Therefore, unsupervised semantic feature learning, i.e., learning without requiring manual annotation effort, is of crucial importance in order to successfully harvest the vast amount of visual data that are available today. In our work we propose to learn image features by training ConvNets to recognize the 2d rotation that is applied to the image that it gets as input. We demonstrate both qualitatively and quantitatively that this apparently simple task actually provides a very powerful supervisory signal for semantic feature learning. We exhaustively evaluate our method in various unsupervised feature learning benchmarks and we exhibit in all of them state-of-the-art performance. Specifically, our results on those benchmarks demonstrate dramatic improvements w.r.t. prior state-of-the-art approaches in unsupervised representation learning and thus significantly close the gap with supervised feature learning. For instance, in PASCAL VOC 2007 detection task our unsupervised pre-trained AlexNet model achieves the state-of-the-art (among unsupervised methods) mAP of 54.4% that is only 2.4 points lower from the supervised case. We get similarly striking results when we transfer our unsupervised learned features on various other tasks, such as ImageNet classification, PASCAL classification, PASCAL segmentation, and CIFAR-10 classification. The code and models of our paper will be published on: https://github.com/gidariss/FeatureLearningRotNet .

In this paper we show how to learn directly from image data (i.e., without resorting to manually-designed features) a general similarity function for comparing image patches, which is a task of fundamental importance for many computer vision problems. To encode such a function, we opt for a CNN-based model that is trained to account for a wide variety of changes in image appearance. To that end, we explore and study multiple neural network architectures, which are specifically adapted to this task. We show that such an approach can significantly outperform the state-of-the-art on several problems and benchmark datasets.

Abstract Drainage basins in many parts of the world are ungauged or poorly gauged, and in some cases existing measurement networks are declining. The problem is compounded by the impacts of human-induced changes to the land surface and climate, occurring at the local, regional and global scales. Predictions of ungauged or poorly gauged basins under these conditions are highly uncertain. The IAHS Decade on Predictions in Ungauged Basins, or PUB, is a new initiative launched by the International Association of Hydrological Sciences (IAHS), aimed at formulating and implementing appropriate science programmes to engage and energize the scientific community, in a coordinated manner, towards achieving major advances in the capacity to make predictions in ungauged basins. The PUB scientific programme focuses on the estimation of predictive uncertainty, and its subsequent reduction, as its central theme. A general hydrological prediction system contains three components: (a) a model that describes the key processes of interest, (b) a set of parameters that represent those landscape properties that govern critical processes, and (c) appropriate meteorological inputs (where needed) that drive the basin response. Each of these three components of the prediction system, is either not known at all, or at best known imperfectly, due to the inherent multi-scale space—time heterogeneity of the hydrological system, especially in ungauged basins. PUB will therefore include a set of targeted scientific programmes that attempt to make inferences about climatic inputs, parameters and model structures from available but inadequate data and process knowledge, at the basin of interest and/or from other similar basins, with robust measures of the uncertainties involved, and their impacts on predictive uncertainty. Through generation of improved understanding, and methods for the efficient quantification of the underlying multi-scale heterogeneity of the basin and its response, PUB will inexorably lead to new, innovative methods for hydrological predictions in ungauged basins in different parts of the world, combined with significant reductions of predictive uncertainty. In this way, PUB will demonstrate the value of data, as well as provide the information needed to make predictions in ungauged basins, and assist in capacity building in the use of new technologies. This paper presents a summary of the science and implementation plan of PUB, with a call to the hydrological community to participate actively in the realization of these goals.

A number of problems can be formulated as prediction on graph-structured data. In this work, we generalize the convolution operator from regular grids to arbitrary graphs while avoiding the spectral domain, which allows us to handle graphs of varying size and connectivity. To move beyond a simple diffusion, filter weights are conditioned on the specific edge labels in the neighborhood of a vertex. Together with the proper choice of graph coarsening, we explore constructing deep neural networks for graph classification. In particular, we demonstrate the generality of our formulation in point cloud classification, where we set the new state of the art, and on a graph classification dataset, where we outperform other deep learning approaches.

We study the plane shear flow of a dense assembly of dissipative disks using discrete simulation and prescribing the pressure and the shear rate. Those shear states are steady and uniform, and become intermittent in the quasistatic regime. In the limit of rigid grains, the shear state is determined by a single dimensionless number, called the inertial number I , which describes the ratio of inertial to pressure forces. Small values of I correspond to the quasistatic critical state of soil mechanics, while large values of I correspond to the fully collisional regime of kinetic theory. When I increases in the intermediate dense flow regime, we measure an approximately linear decrease of the solid fraction from the maximum packing value, and an approximately linear increase of the effective friction coefficient from the static internal friction value. From those dilatancy and friction laws, we deduce the constitutive law for dense granular flows, with a plastic Coulomb term and a viscous Bagnold term. The mechanical characteristics of the grains (restitution, friction, and elasticity) have a small influence in the dense flow regime. Finally, we show that the evolution of the relative velocity fluctuations and of the contact force anisotropy as a function of I provides a simple explanation of the friction law.

The purpose of this paper is twofold. First, we present an alternative model of agglomeration and trade that displays the main features of the recent economic geography literature while allowing for the derivation of analytical results by means of simple algebra. Second, we show how this framework can be used to adopt a forward-looking approach to the dynamics of migration in the process of agglomeration instead of the myopic Marshallian model used so far in this literature.

We commonly refer to state estimation theory in geosciences as data assimilation (DA). This term encompasses the entire sequence of operations that, starting from the observations of a system, and from additional statistical and dynamical information (such as a dynamical evolution model), provides an estimate of its state. DA is standard practice in numerical weather prediction, but its application is becoming widespread in many other areas of climate, atmosphere, ocean, and environment modeling; in all circumstances where one intends to estimate the state of a large dynamical system based on limited information. While the complexity of DA, and of the methods thereof, stands on its interdisciplinary nature across statistics, dynamical systems, and numerical optimization, when applied to geosciences, an additional difficulty arises by the continually increasing sophistication of the environmental models. Thus, in spite of DA being nowadays ubiquitous in geosciences, it has so far remained a topic mostly reserved to experts. We aim this overview article at geoscientists with a background in mathematical and physical modeling, who are interested in the rapid development of DA and its growing domains of application in environmental science, but so far have not delved into its conceptual and methodological complexities. This article is categorized under: Climate Models and Modeling > Knowledge Generation with Models.

We propose a regularized, fast, and robust analytical solution for the Q-ball imaging (QBI) reconstruction of the orientation distribution function (ODF) together with its detailed validation and a discussion on its benefits over the state-of-the-art. Our analytical solution is achieved by modeling the raw high angular resolution diffusion imaging signal with a spherical harmonic basis that incorporates a regularization term based on the Laplace-Beltrami operator defined on the unit sphere. This leads to an elegant mathematical simplification of the Funk-Radon transform which approximates the ODF. We prove a new corollary of the Funk-Hecke theorem to obtain this simplification. Then, we show that the Laplace-Beltrami regularization is theoretically and practically better than Tikhonov regularization. At the cost of slightly reducing angular resolution, the Laplace-Beltrami regularization reduces ODF estimation errors and improves fiber detection while reducing angular error in the ODF maxima detected. Finally, a careful quantitative validation is performed against ground truth from synthetic data and against real data from a biological phantom and a human brain dataset. We show that our technique is also able to recover known fiber crossings in the human brain and provides the practical advantage of being up to 15 times faster than original numerical QBI method.

We propose an object detection system that relies on a multi-region deep convolutional neural network (CNN) that also encodes semantic segmentation-aware features. The resulting CNN-based representation aims at capturing a diverse set of discriminative appearance factors and exhibits localization sensitivity that is essential for accurate object localization. We exploit the above properties of our recognition module by integrating it on an iterative localization mechanism that alternates between scoring a box proposal and refining its location with a deep CNN regression model. Thanks to the efficient use of our modules, we detect objects with very high localization accuracy. On the detection challenges of PASCAL VOC2007 and PASCAL VOC2012 we achieve mAP of 78.2% and 73.9% correspondingly, surpassing any other published work by a significant margin.

With the recent increase in underground urban development as well as for transportation, tunnels need to be driven in increasingly difficult soil conditions. In most cases the ground itself is not stable and face stability is achieved by applying fluid pressure to the tunnel front. The question of determining the retaining pressure to be used has received special consideration in the past because of the concern for safety during construction, and also because of the damage that could be caused to surface structures by the failure of a shallow tunnel. The problem is three-dimensional and can be studied by using the limit state design method. Solutions are available for the case of a circular tunnel in purely cohesive ground, but very little is known of the face stability of a tunnel driven in sandy soils. This Paper approaches this latter problem from the point of view of limit analysis. Both safety against face collapse and blow-out are considered. Three upper bound solutions are derived from the consideration of mechanisms based on the motion of rigid conical blocks. The results are compared with lower bound solutions published in a previous article. A failure criterion is proposed for the tunnel face in the general case of a cohesive and frictional soil, and charts are provided to allow a bracketed estimate of the required retaining pressure. A comparison with centrifuge laboratory tests shows close agreement between the theoretical upper bound solutions and the face pressures at collapse measured experimentally. L'utilisation croissante du sous-sol en site urbain ainsi que le développement des réseaux de transports enterrés conduit à construire des runnels dans des conditions toujours plus difficiles. Dans la plupart des cas le terrain est instable et il est nécessaire d'appliquer une pression de soutènement au niveau du front de taille. Le choix de la pression à utiliser a fait l'objet de nombreuses études au cours des dernières années en raison des problèmes posés pour de tels projets du point de vue de la sécurité en cours de construction et des conséquences d'une rupture sur les structures situées en surface. Il s'agit d'un problème tridimensionnel qui peut être notamment étudié par la méthode de l'analyse limite. Des solutions ont déjà été proposés pour le cas d'un tunnel circulaire en terrain cohérent, mais les connaissances sont encore limitées quant à la stabilité du front de taille d'un tunnel creusé dans des sols sableux. Ce dernier problème est examiné dans le présent article du point de vue de l'analyse limite. On s'intéresse à la fois aux risques de rupture par effondrement et par explosion. L'examen de trois mécanismes basés sur le mouvement rigide de blocs coniques permet d'aboutir à une approche par l'extérieur des conditions de stabilité. Les résultats sont comparés à l'approche par l'intérieur présentée dans un article précédent. L'étude aboutit à la mise en évidence d'un critère de rupture pour le front de taille dans le cas général d'un sol frottant-cohérent. On propose également une série d'abaques permettant d'obtenir un encadrement de la pression de soutènement à utiliser. L'application à des essais en centrifugeuse montre que les bornes supérieures obtenues analytiquement sont très proches des pressions mesurées expérimentalement à la rupture.

Transferring knowledge from a teacher neural network pretrained on the same or a similar task to a student neural network can significantly improve the performance of the student neural network. Existing knowledge transfer approaches match the activations or the corresponding hand-crafted features of the teacher and the student networks. We propose an information-theoretic framework for knowledge transfer which formulates knowledge transfer as maximizing the mutual information between the teacher and the student networks. We compare our method with existing knowledge transfer methods on both knowledge distillation and transfer learning tasks and show that our method consistently outperforms existing methods. We further demonstrate the strength of our method on knowledge transfer across heterogeneous network architectures by transferring knowledge from a convolutional neural network (CNN) to a multi-layer perceptron (MLP) on CIFAR-10. The resulting MLP significantly outperforms the-state-of-the-art methods and it achieves similar performance to the CNN with a single convolutional layer.

In order to drive safely and efficiently on public roads, autonomous vehicles will have to understand the intentions of surrounding vehicles, and adapt their own behavior accordingly. If experienced human drivers are generally good at inferring other vehicles' motion up to a few seconds in the future, most current Advanced Driving Assistance Systems (ADAS) are unable to perform such medium-term forecasts, and are usually limited to high-likelihood situations such as emergency braking. In this article, we present a first step towards consistent trajectory prediction by introducing a long short-term memory (LSTM) neural network, which is capable of accurately predicting future longitudinal and lateral trajectories for vehicles on highway. Unlike previous work focusing on a low number of trajectories collected from a few drivers, our network was trained and validated on the NGSIM US-101 dataset, which contains a total of 800 hours of recorded trajectories in various traffic densities, representing more than 6000 individual drivers.

This paper investigates when patent races will be characterized by vigorous competition and when they will degenerate into a monopoly. Under some conditions, a firm with an arbitrarily small headstart can preempt its rivals. Such 'e-preemption' is shown to depend on whether a firm that is behind in the patent race, as measured by the expected time remaining until discovery, can't 'leapfrog' the competition and become the new leader. An example of an R and D game with random discovery illustrates how e-preemption can occur when leapfrogging is impossible. A multi-stage R and D process allows leapfrogging and thus permits competition. A similar conclusion emerges in a model of a deterministic patent race with imperfect monitoring of rival firms' R and D investment activities.

Within the framework of an extended elastoplastic constitutive model for unsaturated soils (loading-collapse (LC) model), an experimental programme was performed in an osmotically controlled suction triaxial apparatus. The laboratory behaviour of a statically compacted silt was studied, and particular attention was given to the volume changes monitored during shear. Isotropic loading tests confirmed the main features of the LC model related to the effect of suction on volume changes, and allowed a direct determination of the LC curve. Constant σ 3 and a few constant q shear tests were performed in order to study yielding and plastic flow at various increasing suctions, starting from the as-compacted condition. Several yield criteria were considered, depending on the type of test performed. Some similarities between compacted unsaturated soils and natural soft soils were shown, such as the inclined elliptical form of the yield curve, which results from the anisotropic state of stress prevailing during k 0 compaction. Some preconsolidation effects due to increasing suction were identified, and an approximately isotropic suction hardening phenomenon was evidenced, together with a non-associated flow rule. The direction of the plastic strain increment seemed almost independent of the suction, and a hyperbolic plastic potential, similar to that of sand, was found satisfactory. Inclined ellipses were chosen for modelling the yield curves. As for any simple elasto-plastic Cam clay type model applied to overconsolidated soils, the predicted stress—strain curves showed a sudden transition at yield, whereas a much more gradual transition was observed in practice. Volume change prediction appeared satisfactory, showing the validity of the hyperbolic plastic potential. Un programme expérimental, mené à Paide d'un appareil triaxial à succion contrôlée par un dispositif osmotique, a été é1aboré dans un contexte élasto-plastique développé pour les sole non saturés (modèle LC). On a étudié au laboratoire le comportement mécanique d'un limon compacté statiquement, en accordant une importance particulière à la mesure des variations volumiques se produisant pendant le cisaillement. Des essais de compression isotropes ont confirmé les aspects essentiels du modèle LC qui traitent des effets de la succion sur les variations de volume, et ont permis la détermination de la courbe LC. Des essais de cisaillement triaxial à σ 3 constant, et quelques tests à constant ont été réalisés pour étudier la plastification et l'écoulement plastique du matériau à différentes succions, égales on supérieures à la succion de compactage. On a considéré, suivant le type d'essai réalisé, divers critères de plastification. Un certain nombre d'analogies avec les propriétés des sols mous naturels ont été mises en évidence sur les sols compactés non saturés: la forme de la surface de charge est elliptique inclinée, du fait de Panisotropie de état de contrainte appliqué pendant le compactage. Des effets de préconsolidation dus à une succion croissante ont été identifiés, ainsi qu'un écrouissage en succion isotrope et une regle d'écoulement non associée. La direction de l'incrément de déformation plastique s'est ré vélée têtre indépendante de la succion. Comme dans les sables, un potentiel plastique hyperbolique est apparu satisfaisant. On a modélisé les surfaces de charge par des ellipses inclinées. Comme dans les modàles Camclay classiques, les courbes effort-déformation calculées avec le modàle présentent une transition marquée à la plastification, alors que cc passage est progressif dans la réalité Les changements de volume calculés sont apparus satisfaisants, ce qui confirme la validité du potentiel plastique hyperbolique.

Background modeling is an important component of many vision systems. Existing work in the area has mostly addressed scenes that consist of static or quasi-static structures. When the scene exhibits a persistent dynamic behavior in time, such an assumption is violated and detection performance deteriorates. In this paper, we propose a new method for the modeling and subtraction of such scenes. Towards the modeling of the dynamic characteristics, optical flow is computed and utilized as a feature in a higher dimensional space. Inherent ambiguities in the computation of features are addressed by using a data-dependent bandwidth for density estimation using kernels. Extensive experiments demonstrate the utility and performance of the proposed approach.