Archive Ouverte INSA Lyon
Not a member yet
75807 research outputs found
Sort by
Compact Mamba Multi-View for Object Detection
No full textMulti-view image analysis is a key enabler for robust perception when single viewpoints provide incomplete or ambiguous observations. This challenge is particularly pronounced in industrial inspection of transparent materials, where view-dependent optical effects, subtle surface degradations, and annotation noise significantly hinder reliable detection and severity assessment. In this work, we introduce a compact and efficient multi-view fusion architecture tailored to such constraints. Our approach combines shared-weight hierarchical encoders with selective state-space modeling to explicitly exploit cross-view and multi-scale correlations. Multi-View Mamba Blocks (MVMB) perform adaptive fusion at each feature level by coupling Mamba-based selective state-space layers with FiLM-driven cross-view conditioning, while a Global State-Space Fusion Block enforces long-range coherence across all views and resolutions. Task-specific decoding heads query the resulting global representation via cross-attention to jointly predict object localization and ordinal wear severity. The model is trained using a unified multi-task objective that integrates geometric regression, ordinal classification, cross-view consistency, feature alignment, and sequential smoothness. Extensive experiments on a challenging multi-view glass container inspection dataset demonstrate improved robustness, consistency, and scalability compared to strong baselines. To promote reproducibility and future research, we publicly release the proposed dataset at: https://datasets.liris.cnrs.fr/mvep-version1
Electroporation of spheroids using an electric field gradient: a tool to study intensity-dependent permeabilization
No full textInternational audienceElectroporation (EPN) is the process by which cell membranes become transiently or permanently permeable upon exposure to pulsed electric fields of suitable intensity and duration. Depending on the pulse parameters, permeabilization can be reversible or irreversible, enabling a wide range of biomedical applications. To improve our understanding of EPN effects on tissues and select efficient treatments and parameters, relevant in vitro tumour models are required. Three-dimensional (3D) cell spheroids have emerged as valuable systems, as they more accurately replicate tumour microenvironment and cell-to-cell interactions than conventional 2D cultures. In this work, we present a new microdevice designed for the culture and gradual electroporation of a population of several hundred uniformly sized spheroids, allowing the systematic study of electroporation over a wide range of electric field intensities within a single experiment. Gradual permeabilization of HT-29 colorectal cancer cell spheroids was performed using a standard electrochemotherapy protocol, and electroporation efficiency was assessed by analysing propidium iodide (PI) uptake. Spheroids were treated with electric fields ranging from 800 V cm−1 to 3400 V cm−1. In toto analysis of PI distribution within spheroids by confocal microscopy revealed highly heterogeneous permeabilization patterns throughout the spheroid volume, for all intensities investigated, even at the highest one of 3400 V cm−1. This study introduces a robust 3D in vitro assay for the systematic evaluation of electroporation-based treatments, providing new insights into the influence of electric field heterogeneity, electrical protocol, and estimation of molecular uptake in the spheroid volume
A localisation phase transition for the catalytic branching random walk
No full textWe show the existence of a phase transition between a localisation and a non-localisation regime for a branching random walk with a catalyst at the origin. More precisely, we consider a continuous-time branching random walk that jumps at rate one, with simple random walk jumps on , and that branches (with binary branching) at rate everywhere, except at the origin, where it branches at rate . We show that, if is large enough, then the occupation measure of the branching random walk localises (i.e.~when normalised by the total number of particles, it converges almost surely without spatial renormalisation), whereas, if is close enough to , then the occupation measure delocalises, in the sense that the proportion of particles in any finite given set converges almost surely to zero. The case (when branching only occurs at the origin) has been extensively studied in the literature and a transition between localisation and non-localisation was also exhibited in this case. Interestingly, the transition that we observe, conjecture, and partially prove in this paper occurs at the same threshold as in the case~.One of the strengths of our result is that, in the localisation regime, we are able to prove convergence of the occupation measure, whilst existing results in the case give convergence of moments instead
Model selection for extremal dependence structures using deep learning: Application to environmental data
No full textAlthough the CLIC-based model selection approach is widely used to identify spatial extreme models, the complexity of the associated statistical inference limits the reliability of this criterion. In addition, the strong spatial dependence in small or moderate regions may lead to substantial overlap among the spatial extremes models. This potential overlap increases the risk of model misidentification. In this paper, we exploit the ability of Convolutional Neural Networks (CNNs) to extract spatial patterns in order to develop a CNN-based model selection framework. The proposed approach evaluates how well the dependence structure observed in the data matches the dependence patterns implied by competing models. Two identification strategies are considered. In the first strategy, both the max-stable model and its associated covariance function are identified simultaneously by a single CNN in a one-step procedure. In the second strategy, model identification is performed hierarchically. First, a CNN identifies the class of max-stable model, and then additional CNNs are trained for each model to determine the corresponding covariance function. The performance of the two strategies is evaluated through an extensive simulation study designed to reproduce the spatial dependence structure of 2-m air temperature data over Iraq, where strong dependence and model overlap are observed. The results demonstrate that the proposed CNN-based approach provides an effective alternative for model selection in spatial extremes
Quantization-aware training: a tradeoff between training and fine-tuning for domain-specific language models
No full textInternational audienceQuantization is a widely adopted technique to reduce memory footprint and computational cost in neural networks. While quantizing pre-trained models is effective, retraining is often required for extreme quantization formats. Fine-tuning, on the other hand, enables the adaptation of general-purpose models to specific domains, but quantization can significantly degrade their performance. In this work, we investigate the training cost of finetuned and quantized language models. By formalizing the computational trade-off between domain adaptation and fine-tuning, we demonstrate that domain-specialized checkpoints exhibit greater robustness to quantization noise. Our findings establish a viable blueprint for deploying high-performance biomedical NLP models in resource-constrained, edge environments
Séries génératrices de matroïde
No full textGiven a finite set - for instance, the set of edges in a finite multigraph - it is customary to study the poset of its subsets and, in particular, its incidence algebra, which forms a subalgebra of the algebra of upper triangular matrices. As a vector space, this algebra is spanned by pairs of subsets satisfying . Notably, there is a bijection between these pairs and the minors of a matroid (or graph) defined on ; here, the elements (edges) in are interpreted as being deleted, while those in are contracted, so that the support of the resulting minor is given by .This is why our incidence algebra is highly useful in matroid theory: it can be employed just like formal power series, notably by supporting operations such as differentiation and substitution. This approach is closely linked to the axiomatic characterizations of matroids and to a careful study of the greedy algorithm. Consequently, it should enable us to tackle, at the very least, every exact enumeration problem related to matroids and oriented matroids (including Tutte polynomials). To illustrate this idea, we provide concise formulations and proofs for many classical and new results in the theories of graphs, matroids, and oriented matroids, emphasizing the benefits of studying a matroid together with its minors as a single object - what we call a matroid power series
Digital 3D Technologies for Preserving and Transmitting Cultural heritage and Historical Memories
No full textInternational audienceIn an era marked by both environmental transformations and the resurgence of conflicts, the preservation of cultural and historical memory faces unprecedented challenges. Whether submerged by dam reservoirs or obscured by the passage of time, landscapes and cultural heritages risk fading from collective memory. Yet, these elements are crucial for educating current and future generations about the impacts of human activity or conflict on local populations. This communication explores how 3D digital technologies, particularly 3D Geographic Information Systems (GIS) and geovisualization, can restore these threatened memories and make them accessible to a broad audience. Using two case studies, we will demonstrate the potential of these tools to reconstruct, archive, and disseminate lost or endangered heritage.First, we will examine the Gorges de la Loire valley in France, where the construction of reservoir dams submerged landscapes and cultural sites . By combining 3D GIS with historical databases, we created a "backup copy" of the flooded valley, enabling local populations to re-appropriate their submerged heritage. This approach not only preserves the memory of lost landscapes but also provides a model for safeguarding other sites threatened by environmental changes.Second, we turn to Saint-Étienne during World War II, where events and local community memories risk being forgotten . Here, 3D GIS and virtual reality tools were used to archive and visualize the city’s wartime experiences, from resistance, occupation and collaboration acts to bombed neighborhoods.By integrating maps, 3D models, and immersive videos, we developed a multidimensional archive that engages diverse audiences – from researchers to students and local inhabitants – in exploring the past.Our methodology highlights three key contributions:1.Technological Innovation: The use of 3D GIS and geovisualization to create dynamic, interactive representations of lost or threatened heritage.2.Cultural Preservation: The development of digital archives that serve as memory backups.3.Public Engagement: The adaptation of these tools for educational and commemorative purposes, fostering a deeper understanding of historical and environmental impacts
STORM: Slot-based Task-aware Object-centric Representation for robotic Manipulation
No full textVisual foundation models provide strong perceptual features for robotics, but their dense representations lack explicit object-level structure, limiting robustness and contractility in manipulation tasks. We propose STORM (Slot-based Task-aware Object-centric Representation for robotic Manipulation), a lightweight object-centric adaptation module that augments frozen visual foundation models with a small set of semantic-aware slots for robotic manipulation. Rather than retraining large backbones, STORM employs a multi-phase training strategy: object-centric slots are first stabilized through visual–semantic pretraining using language embeddings, then jointly adapted with a downstream manipulation policy. This staged learning prevents degenerate slot formation and preserves semantic consistency while aligning perception with task objectives. Experiments on object discovery benchmarks and simulated manipulation tasks show that STORM improves generalization to visual distractors, and control performance compared to directly using frozen foundation model features or training object-centric representations end-to-end. Our results highlight multi-phase adaptation as an efficient mechanism for transforming generic foundation model features into task-aware object-centric representations for robotic control