Université Gustave Eiffel

Deep residual networks were shown to be able to scale up to thousands of layers and still have improving performance. However, each fraction of a percent of improved accuracy costs nearly doubling the number of layers, and so training very deep residual networks has a problem of diminishing feature reuse, which makes these networks very slow to train. To tackle these problems, in this paper we conduct a detailed experimental study on the architecture of ResNet blocks, based on which we propose a novel architecture where we decrease depth and increase width of residual networks. We call the resulting network structures wide residual networks (WRNs) and show that these are far superior over their commonly used thin and very deep counterparts. For example, we demonstrate that even a simple 16-layer-deep wide residual network outperforms in accuracy and efficiency all previous deep residual networks, including thousand-layer-deep networks, achieving new state-of-the-art results on CIFAR, SVHN, COCO, and significant improvements on ImageNet. Our code and models are available at this https URL

Scene labeling consists of labeling each pixel in an image with the category of the object it belongs to. We propose a method that uses a multiscale convolutional network trained from raw pixels to extract dense feature vectors that encode regions of multiple sizes centered on each pixel. The method alleviates the need for engineered features, and produces a powerful representation that captures texture, shape, and contextual information. We report results using multiple postprocessing methods to produce the final labeling. Among those, we propose a technique to automatically retrieve, from a pool of segmentation components, an optimal set of components that best explain the scene; these components are arbitrary, for example, they can be taken from a segmentation tree or from any family of oversegmentations. The system yields record accuracies on the SIFT Flow dataset (33 classes) and the Barcelona dataset (170 classes) and near-record accuracy on Stanford background dataset (eight classes), while being an order of magnitude faster than competing approaches, producing a $(320\times 240)$ image labeling in less than a second, including feature extraction.

Advanced oxidation processes (AOPs) constitute important, promising, efficient, and environmental-friendly methods developed to principally remove persistent organic pollutants (POPs) from waters and wastewaters. Generally, AOPs are based on the in situ generation of a powerful oxidizing agent, such as hydroxyl radicals (•OH), obtained at a sufficient concentration to effectively decontaminate waters. This critical review presents a precise and overall description of the recent literature (period 1990–2012) concerning the main types of AOPs, based on chemical, photochemical, sonochemical, and electrochemical reactions. The principles, performances, advantages, drawbacks, and applications of these AOPs to the degradation and destruction of POPs in aquatic media and to the treatment of waters and waste waters have been reported and compared.

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.

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.

The 12th generation of the International Geomagnetic Reference Field (IGRF) was adopted in December 2014 by the Working Group V-MOD appointed by the International Association of Geomagnetism and Aeronomy (IAGA). It updates the previous IGRF generation with a definitive main field model for epoch 2010.0, a main field model for epoch 2015.0, and a linear annual predictive secular variation model for 2015.0-2020.0. Here, we present the equations defining the IGRF model, provide the spherical harmonic coefficients, and provide maps of the magnetic declination, inclination, and total intensity for epoch 2015.0 and their predicted rates of change for 2015.0-2020.0. We also update the magnetic pole positions and discuss briefly the latest changes and possible future trends of the Earth’s magnetic field.

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 In December 2019, the International Association of Geomagnetism and Aeronomy (IAGA) Division V Working Group (V-MOD) adopted the thirteenth generation of the International Geomagnetic Reference Field (IGRF). This IGRF updates the previous generation with a definitive main field model for epoch 2015.0, a main field model for epoch 2020.0, and a predictive linear secular variation for 2020.0 to 2025.0. This letter provides the equations defining the IGRF, the spherical harmonic coefficients for this thirteenth generation model, maps of magnetic declination, inclination and total field intensity for the epoch 2020.0, and maps of their predicted rate of change for the 2020.0 to 2025.0 time period.

We report on the population properties of compact binary mergers inferred from gravitational-wave observations of these systems during the first three LIGO-Virgo observing runs. The Gravitational-Wave Transient Catalog 3 (GWTC-3) contains signals consistent with three classes of binary mergers: binary black hole, binary neutron star, and neutron star–black hole mergers. We infer the binary neutron star merger rate to be between 10 and <a:math xmlns:a="http://www.w3.org/1998/Math/MathML" display="inline"><a:mrow><a:mn>1700</a:mn><a:mtext> </a:mtext><a:mtext> </a:mtext><a:msup><a:mrow><a:mi>Gpc</a:mi></a:mrow><a:mrow><a:mo>−</a:mo><a:mn>3</a:mn></a:mrow></a:msup><a:mtext> </a:mtext><a:msup><a:mrow><a:mi>yr</a:mi></a:mrow><a:mrow><a:mo>−</a:mo><a:mn>1</a:mn></a:mrow></a:msup></a:mrow></a:math> and the neutron star–black hole merger rate to be between 7.8 and <c:math xmlns:c="http://www.w3.org/1998/Math/MathML" display="inline"><c:mrow><c:mn>140</c:mn><c:mtext> </c:mtext><c:mtext> </c:mtext><c:msup><c:mrow><c:mi>Gpc</c:mi></c:mrow><c:mrow><c:mo>−</c:mo><c:mn>3</c:mn></c:mrow></c:msup><c:mtext> </c:mtext><c:msup><c:mrow><c:mi>yr</c:mi></c:mrow><c:mrow><c:mo>−</c:mo><c:mn>1</c:mn></c:mrow></c:msup></c:mrow></c:math>, assuming a constant rate density in the comoving frame and taking the union of 90% credible intervals for methods used in this work. We infer the binary black hole merger rate, allowing for evolution with redshift, to be between 17.9 and <e:math xmlns:e="http://www.w3.org/1998/Math/MathML" display="inline"><e:mrow><e:mn>44</e:mn><e:mtext> </e:mtext><e:mtext> </e:mtext><e:msup><e:mrow><e:mi>Gpc</e:mi></e:mrow><e:mrow><e:mo>−</e:mo><e:mn>3</e:mn></e:mrow></e:msup><e:mtext> </e:mtext><e:msup><e:mrow><e:mi>yr</e:mi></e:mrow><e:mrow><e:mo>−</e:mo><e:mn>1</e:mn></e:mrow></e:msup></e:mrow></e:math> at a fiducial redshift (<g:math xmlns:g="http://www.w3.org/1998/Math/MathML" display="inline"><g:mi>z</g:mi><g:mo>=</g:mo><g:mn>0.2</g:mn></g:math>). The rate of binary black hole mergers is observed to increase with redshift at a rate proportional to <i:math xmlns:i="http://www.w3.org/1998/Math/MathML" display="inline"><i:mo stretchy="false">(</i:mo><i:mn>1</i:mn><i:mo>+</i:mo><i:mi>z</i:mi><i:msup><i:mo stretchy="false">)</i:mo><i:mi>κ</i:mi></i:msup></i:math> with <m:math xmlns:m="http://www.w3.org/1998/Math/MathML" display="inline"><m:mi>κ</m:mi><m:mo>=</m:mo><m:mn>2.</m:mn><m:msubsup><m:mn>9</m:mn><m:mrow><m:mo>−</m:mo><m:mn>1.8</m:mn></m:mrow><m:mrow><m:mo>+</m:mo><m:mn>1.7</m:mn></m:mrow></m:msubsup></m:math> for <o:math xmlns:o="http://www.w3.org/1998/Math/MathML" display="inline"><o:mi>z</o:mi><o:mo>≲</o:mo><o:mn>1</o:mn></o:math>. Using both binary neutron star and neutron star–black hole binaries, we obtain a broad, relatively flat neutron star mass distribution extending from <q:math xmlns:q="http://www.w3.org/1998/Math/MathML" display="inline"><q:msubsup><q:mn>1.2</q:mn><q:mrow><q:mo>−</q:mo><q:mn>0.2</q:mn></q:mrow><q:mrow><q:mo>+</q:mo><q:mn>0.1</q:mn></q:mrow></q:msubsup></q:math> to <s:math xmlns:s="http://www.w3.org/1998/Math/MathML" display="inline"><s:msubsup><s:mn>2.0</s:mn><s:mrow><s:mo>−</s:mo><s:mn>0.3</s:mn></s:mrow><s:mrow><s:mo>+</s:mo><s:mn>0.3</s:mn></s:mrow></s:msubsup><s:msub><s:mi>M</s:mi><s:mo stretchy="false">⊙</s:mo></s:msub></s:math>. We confidently determine that the merger rate as a function of mass sharply declines after the expected maximum neutron star mass, but cannot yet confirm or rule out the existence of a lower mass gap between neutron stars and black holes. We also find the binary black hole mass distribution has localized over- and underdensities relative to a power-law distribution, with peaks emerging at chirp masses of <v:math xmlns:v="http://www.w3.org/1998/Math/MathML" display="inline"><v:msubsup><v:mn>8.3</v:mn><v:mrow><v:mo>−</v:mo><v:mn>0.5</v:mn></v:mrow><v:mrow><v:mo>+</v:mo><v:mn>0.3</v:mn></v:mrow></v:msubsup></v:math> and <x:math xmlns:x="http://www.w3.org/1998/Math/MathML" display="inline"><x:msubsup><x:mn>27.9</x:mn><x:mrow><x:mo>−</x:mo><x:mn>1.8</x:mn></x:mrow><x:mrow><x:mo>+</x:mo><x:mn>1.9</x:mn></x:mrow></x:msubsup><x:msub><x:mi>M</x:mi><x:mo stretchy="false">⊙</x:mo></x:msub></x:math>. While we continue to find that the mass distribution of a binary’s more massive component strongly decreases as a function of primary mass, we observe no evidence of a strongly suppressed merger rate above approximately <ab:math xmlns:ab="http://www.w3.org/1998/Math/MathML" display="inline"><ab:mn>60</ab:mn><ab:msub><ab:mi>M</ab:mi><ab:mo stretchy="false">⊙</ab:mo></ab:msub></ab:math>, which would indicate the presence of a upper mass gap. Observed black hole spins are small, with half of spin magnitudes below <db:math xmlns:db="http://www.w3.org/1998/Math/MathML" display="inline"><db:msub><db:mi>χ</db:mi><db:mi>i</db:mi></db:msub><db:mo>≈</db:mo><db:mn>0.25</db:mn></db:math>. While the majority of spins are preferentially aligned with the orbital angular momentum, we infer evidence of antialigned spins among the binary population. We observe an increase in spin magnitude for systems with more unequal-mass ratio. We also observe evidence of misalignment of spins relative to the orbital angular momentum. Published by the American Physical Society 2023

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.

cubic meters of rock and glacier ice collapsed from the steep north face of Ronti Peak. The rock and ice avalanche rapidly transformed into an extraordinarily large and mobile debris flow that transported boulders greater than 20 meters in diameter and scoured the valley walls up to 220 meters above the valley floor. The intersection of the hazard cascade with downvalley infrastructure resulted in a disaster, which highlights key questions about adequate monitoring and sustainable development in the Himalaya as well as other remote, high-mountain environments.

Abstract Geophysics provides a multidimensional suite of investigative methods that are transforming our ability to see into the very fabric of the subsurface environment, and monitor the dynamics of its fluids and the biogeochemical reactions that occur within it. Here we document how geophysical methods have emerged as valuable tools for investigating shallow subsurface processes over the past two decades and offer a vision for future developments relevant to hydrology and also ecosystem science. The field of “hydrogeophysics” arose in the late 1990s, prompted, in part, by the wealth of studies on stochastic subsurface hydrology that argued for better field‐based investigative techniques. These new hydrogeophysical approaches benefited from the emergence of practical and robust data inversion techniques, in many cases with a view to quantify shallow subsurface heterogeneity and the associated dynamics of subsurface fluids. Furthermore, the need for quantitative characterization stimulated a wealth of new investigations into petrophysical relationships that link hydrologically relevant properties to measurable geophysical parameters. Development of time‐lapse approaches provided a new suite of tools for hydrological investigation, enhanced further with the realization that some geophysical properties may be sensitive to biogeochemical transformations in the subsurface environment, thus opening up the new field of “biogeophysics.” Early hydrogeophysical studies often concentrated on relatively small “plot‐scale” experiments. More recently, however, the translation to larger‐scale characterization has been the focus of a number of studies. Geophysical technologies continue to develop, driven, in part, by the increasing need to understand and quantify key processes controlling sustainable water resources and ecosystem services.

Coffee is a valuable beverage crop due to its characteristic flavor, aroma, and the stimulating effects of caffeine. We generated a high-quality draft genome of the species Coffea canephora, which displays a conserved chromosomal gene order among asterid angiosperms. Although it shows no sign of the whole-genome triplication identified in Solanaceae species such as tomato, the genome includes several species-specific gene family expansions, among them N-methyltransferases (NMTs) involved in caffeine production, defense-related genes, and alkaloid and flavonoid enzymes involved in secondary compound synthesis. Comparative analyses of caffeine NMTs demonstrate that these genes expanded through sequential tandem duplications independently of genes from cacao and tea, suggesting that caffeine in eudicots is of polyphyletic origin.

We report the observation of gravitational waves from two compact binary coalescences in LIGO's and Virgo's third observing run with properties consistent with neutron star-black hole (NSBH) binaries. The two events are named GW200105_162426 and GW200115_042309, abbreviated as GW200105 and GW200115; the first was observed by LIGO Livingston and Virgo and the second by all three LIGO-Virgo detectors. The source of GW200105 has component masses, whereas the source of GW200115 has component masses and (all measurements quoted at the 90% credible level). The probability that the secondary's mass is below the maximal mass of a neutron star is 89%-96% and 87%-98%, respectively, for GW200105 and GW200115, with the ranges arising from different astrophysical assumptions. The source luminosity distances are and, respectively. The magnitude of the primary spin of GW200105 is less than 0.23 at the 90% credible level, and its orientation is unconstrained. For GW200115, the primary spin has a negative spin projection onto the orbital angular momentum at 88% probability. We are unable to constrain the spin or tidal deformation of the secondary component for either event. We infer an NSBH merger rate density of when assuming that GW200105 and GW200115 are representative of the NSBH population or under the assumption of a broader distribution of component masses. © 2021. The Author(s). Published by the American Astronomical Society.

This article addresses educational challenges posed by the future of work, examining “21st century skills”, their conception, assessment, and valorization. It focuses in particular on key soft skill competencies known as the “4Cs”: creativity, critical thinking, collaboration, and communication. In a section on each C, we provide an overview of assessment at the level of individual performance, before focusing on the less common assessment of systemic support for the development of the 4Cs that can be measured at the institutional level (i.e., in schools, universities, professional training programs, etc.). We then present the process of official assessment and certification known as “labelization”, suggesting it as a solution both for establishing a publicly trusted assessment of the 4Cs and for promoting their cultural valorization. Next, two variations of the “International Institute for Competency Development’s 21st Century Skills Framework” are presented. The first of these comprehensive systems allows for the assessment and labelization of the extent to which development of the 4Cs is supported by a formal educational program or institution. The second assesses informal educational or training experiences, such as playing a game. We discuss the overlap between the 4Cs and the challenges of teaching and institutionalizing them, both of which may be assisted by adopting a dynamic interactionist model of the 4Cs—playfully entitled “Crea-Critical-Collab-ication”—for pedagogical and policy-promotion purposes. We conclude by briefly discussing opportunities presented by future research and new technologies such as artificial intelligence and virtual reality.

Surface wave methods gained in the past decades a primary role in many seismic projects. Specifically, they are often used to retrieve a 1D shear wave velocity model or to estimate the VS,30 at a site. The complexity of the interpretation process and the variety of possible approaches to surface wave analysis make it very hard to set a fixed standard to assure quality and reliability of the results. The present guidelines provide practical information on the acquisition and analysis of surface wave data by giving some basic principles and specific suggestions related to the most common situations. They are primarily targeted to non-expert users approaching surface wave testing, but can be useful to specialists in the field as a general reference. The guidelines are based on the experience gained within the InterPACIFIC project and on the expertise of the participants in acquisition and analysis of surface wave data.

Under consideration is the large body of signal recovery problems that can be formulated as the problem of minimizing the sum of two (not necessarily smooth) lower semicontinuous convex functions in a real Hilbert space. This generic problem is analyzed and a decomposition method is proposed to solve it. The convergence of the method, which is based on the Douglas-Rachford algorithm for monotone operator-splitting, is obtained under general conditions. Applications to non-Gaussian image denoising in a tight frame are also demonstrated.

Recent experimental observations of the onset of calcium carbonate (CaCO3) mineralization suggest the emergence of a population of clusters that are stable rather than unstable as predicted by classical nucleation theory. This study uses molecular dynamics simulations to probe the structure, dynamics, and energetics of hydrated CaCO3 clusters and lattice gas simulations to explore the behavior of cluster populations before nucleation. Our results predict formation of a dense liquid phase through liquid-liquid separation within the concentration range in which clusters are observed. Coalescence and solidification of nanoscale droplets results in formation of a solid phase, the structure of which is consistent with amorphous CaCO3. The presence of a liquid-liquid binodal enables a diverse set of experimental observations to be reconciled within the context of established phase-separation mechanisms.

This paper proposes a new algorithm for automatic crack detection from 2D pavement images. It strongly relies on the localization of minimal paths within each image, a path being a series of neighboring pixels and its score being the sum of their intensities. The originality of the approach stems from the proposed way to select a set of minimal paths and the two postprocessing steps introduced to improve the quality of the detection. Such an approach is a natural way to take account of both the photometric and geometric characteristics of pavement images. An intensive validation is performed on both synthetic and real images (from five different acquisition systems), with comparisons to five existing methods. The proposed algorithm provides very robust and precise results in a wide range of situations, in a fully unsupervised manner, which is beyond the current state of the art.

Visual understanding is often based on measuring similarity between observations. Learning similarities specific to a certain perception task from a set of examples has been shown advantageous in various computer vision and pattern recognition problems. In many important applications, the data that one needs to compare come from different representations or modalities, and the similarity between such data operates on objects that may have different and often incommensurable structure and dimensionality. In this paper, we propose a framework for supervised similarity learning based on embedding the input data from two arbitrary spaces into the Hamming space. The mapping is expressed as a binary classification problem with positive and negative examples, and can be efficiently learned using boosting algorithms. The utility and efficiency of such a generic approach is demonstrated on several challenging applications including cross-representation shape retrieval and alignment of multi-modal medical images. 1.