Search

Robots in human-centered environments require accurate scene understanding to perform high-level tasks effectively. This understanding can be achieved through instance-aware semantic mapping, which involves reconstructing elements at the level of individual instances. Neural networks, the de facto solution for scene understanding, still face limitations such as overconfident incorrect predictions with out-of-distribution objects or generating inaccurate masks.Placing excessive reliance on these predictions makes the reconstruction susceptible to errors, reducing the robustness of the resulting maps and hampering robot operation. In this work, we propose Voxeland, a probabilistic framework for incrementally building instance-aware semantic maps. Inspired by the Theory of Evidence, Voxeland treats neural network predictions as subjective opinions regarding map instances at both geometric and semantic levels. These opinions are aggregated over time to form evidences, which are formalized through a probabilistic model. This enables us to quantify uncertainty in the reconstruction process, facilitating the identification of map areas requiring improvement (e.g. reobservation or reclassification). As one strategy to exploit this, we incorporate a Large Vision-Language Model (LVLM) to perform semantic level disambiguation for instances with high uncertainty. Results from the standard benchmarking on the publicly available SceneNN dataset demonstrate that Voxeland outperforms state-of-the-art methods, highlighting the benefits of incorporating and leveraging both instance- and semantic-level uncertainties to enhance reconstruction robustness. This is further validated through qualitative experiments conducted on the real-world ScanNet dataset.

The growing demand for accurate and equitable AI models in digital dermatology faces a significant challenge: the lack of diverse, high-quality labeled data. In this work, we investigate the potential of domain-specific foundation models for dermatology in addressing this challenge. We utilize self-supervised learning (SSL) techniques to pre-train models on a dataset of over 240,000 dermatological images from public and private collections. Our study considers several SSL methods and compares the resulting foundation models against domain-agnostic models like those pre-trained on ImageNet and state-of-the-art models such as MONET across 12 downstream tasks. Unlike previous research, we emphasize the development of smaller models that are more suitable for resource-limited clinical settings, facilitating easier adaptation to a broad range of use cases. Results show that models pre-trained in this work not only outperform general-purpose models but also approach the performance of models 50 times larger on clinically relevant diagnostic tasks. To promote further research in this direction, we publicly release both the training code and the foundation models, which can benefit clinicians in dermatological applications., Comment: Findings paper presented at Machine Learning for Health (ML4H) symposium 2024, December 15-16, 2024, Vancouver, Canada, 11 pages

Africa faces a huge shortage of dermatologists, with less than one per million people. This is in stark contrast to the high demand for dermatologic care, with 80% of the paediatric population suffering from largely untreated skin conditions. The integration of AI into healthcare sparks significant hope for treatment accessibility, especially through the development of AI-supported teledermatology. Current AI models are predominantly trained on white-skinned patients and do not generalize well enough to pigmented patients. The PASSION project aims to address this issue by collecting images of skin diseases in Sub-Saharan countries with the aim of open-sourcing this data. This dataset is the first of its kind, consisting of 1,653 patients for a total of 4,901 images. The images are representative of telemedicine settings and encompass the most common paediatric conditions: eczema, fungals, scabies, and impetigo. We also provide a baseline machine learning model trained on the dataset and a detailed performance analysis for the subpopulations represented in the dataset. The project website can be found at https://passionderm.github.io/., Comment: MICCAI 2024

For each prime number $\ell$ and for each imaginary quadratic order of class number one or two, we determine all the possible $\ell$-adic Galois representations that occur for any elliptic curve with complex multiplication by such an order over its minimal field of definition, and then we determine all the isomorphism classes of elliptic curves that have a prescribed $\ell$-adic Galois representation., Comment: 38 pages; this version fixes some minor typos and includes a link in the introduction to a Github repository with code used in the paper

Neutrino event generators make use of intranuclear cascade models (INCs), to predict the kinematics of hadrons produced in neutrino-nucleus interactions. We perform a consistent comparison of different INCs, by using the same set of events as input to the NEUT, NuWro, Achilles and INCL INCs. The inputs correspond to calculations of the fully differential single-proton knockout cross section, either in the distorted-wave impulse approximation (DWIA) or plane-wave impulse approximation (PWIA), both including realistic nuclear hole spectral functions. We compare the INC results to DWIA calculations with an optical potential, used extensively in the analysis of (e,e'p) experiments. We point out a systematic discrepancy between both approaches. We apply the INC results to recent MicroBooNE data. We assess the influence of the choice of spectral function, finding that large variations in realistic spectral functions are indistinguishable with present data. The data is underpredicted, with strength missing in the region where two-nucleon knockout and resonance production contribute. However, the data is underpredicted also in regions of low transverse missing momentum, where one-nucleon knockout dominates. The inclusion of the interference with two-body currents could lead to additional strength in this region., Comment: 26 pages including appendices, comments welcome

Neutrino-induced single-pion production constitutes an essential interaction channel in modern neutrino oscillation experiments, with its products building up a significant fraction of the observable hadronic final states. Frameworks of oscillation analyses strongly rely on Monte Carlo neutrino event generators, which provide theoretical predictions of neutrino interactions on nuclear targets. Thus, it is crucial to integrate state-of-the-art single-pion production models with Monte Carlo simulations to prepare for the upcoming systematics-dominated landscape of neutrino measurements. In this work, we present the implementation of the Ghent Hybrid model for neutrino-induced single-pion production in the NuWro Monte Carlo event generator. The interaction dynamics includes coherently-added contributions from nucleon resonances and a non-resonant background, merged into the pythia branching predictions in the deep-inelastic regime, as instrumented by NuWro. This neutrino-nucleon interaction model is fully incorporated into the nuclear framework of the generator, allowing it to account for the influence of both initial- and final-state nuclear medium effects. We compare the predictions of this integrated implementation with recent pion production data from accelerator-based neutrino experiments. The results of the novel model show improved agreement of the generator predictions with the data and point to the significance of the refined treatment of the description of pion-production processes beyond the $\Delta$ region., Comment: 27 pages, 16 figures, accepted version in JHEP

Out-Of-Distribution (OOD) detection is critical to deploy deep learning models in safety-critical applications. However, the inherent hierarchical concept structure of visual data, which is instrumental to OOD detection, is often poorly captured by conventional methods based on Euclidean geometry. This work proposes a metric framework that leverages the strengths of Hyperbolic geometry for OOD detection. Inspired by previous works that refine the decision boundary for OOD data with synthetic outliers, we extend this method to Hyperbolic space. Interestingly, we find that synthetic outliers do not benefit OOD detection in Hyperbolic space as they do in Euclidean space. Furthermore we explore the relationship between OOD detection performance and Hyperbolic embedding dimension, addressing practical concerns in resource-constrained environments. Extensive experiments show that our framework improves the FPR95 for OOD detection from 22\% to 15\% and from 49% to 28% on CIFAR-10 and CIFAR-100 respectively compared to Euclidean methods., Comment: European Conference on Computer Vision ECCV 2024 BEW Workshop

Theoretical predictions for lepton-induced single-pion production (SPP) on $^{12}$C are revisited in order to assess the effect of different treatments of the current operator. On one hand we have the asymptotic approximation, which consists in replacing the particle four-vectors that enter in the operator by their asymptotic values, i.e., their values out of the nucleus. On the other hand we have the full calculation, which is a more accurate approach to the problem. We also compare with results in which the final nucleon is described by a relativistic plane wave, to rate the effect of the nucleon distortion. The study is performed for several lepton kinematics, reproducing inclusive and semi-inclusive cross sections belonging to the low-$Q^2$ region (between 0.05 and 1 GeV$^2$), which is of special interest in charged-current (CC) neutrino-nucleus 1$\pi$ production. Inclusive electron results are compared with experimental data. We find non-trivial corrections comparable in size with the effect of the nucleon distortion, namely, corrections up to 6\%, either increasing or diminishing the asymptotic prediction, and a shift of the distributions towards higher energy transfer. For the semi-inclusive cross sections, we observe the correction to be prominent mainly at low values of the outgoing nucleon kinetic energy. Finally, for CC neutrino-induced 1$\pi^+$ production, we find a reduction at low-$Q^2$ with respect to both the plane-wave approach and the asymptotic case.

We present a novel 3D odometry method that recovers the full motion of a vehicle only from a Doppler-capable range sensor. It leverages the radial velocities measured from the scene, estimating the sensor's velocity from a single scan. The vehicle's 3D motion, defined by its linear and angular velocities, is calculated taking into consideration its kinematic model which provides a constraint between the velocity measured at the sensor frame and the vehicle frame. Experiments carried out prove the viability of our single-sensor method compared to mounting an additional IMU. Our method provides the translation of the sensor, which cannot be reliably determined from an IMU, as well as its rotation. Its short-term accuracy and fast operation (~5ms) make it a proper candidate to supply the initialization to more complex localization algorithms or mapping pipelines. Not only does it reduce the error of the mapper, but it does so at a comparable level of accuracy as an IMU would. All without the need to mount and calibrate an extra sensor on the vehicle., Comment: This work has been submitted to the IEEE for possible publication

Benchmark datasets for digital dermatology unwittingly contain inaccuracies that reduce trust in model performance estimates. We propose a resource-efficient data-cleaning protocol to identify issues that escaped previous curation. The protocol leverages an existing algorithmic cleaning strategy and is followed by a confirmation process terminated by an intuitive stopping criterion. Based on confirmation by multiple dermatologists, we remove irrelevant samples and near duplicates and estimate the percentage of label errors in six dermatology image datasets for model evaluation promoted by the International Skin Imaging Collaboration. Along with this paper, we publish revised file lists for each dataset which should be used for model evaluation. Our work paves the way for more trustworthy performance assessment in digital dermatology., Comment: Link to the revised file lists: https://github.com/Digital-Dermatology/SelfClean-Revised-Benchmarks

We present a detailed study of the contribution from two-body currents to the one-nucleon knockout process induced by electromagnetic interaction. The framework is a relativistic mean-field model (RMF) in which bound and scattering nucleons are consistently described as solutions of Dirac equation with potentials. We show results obtained with the most general expression of the two-body operator, in which the intermediate nucleons are described by relativistic mean-field bound states; then, we propose two approximations consisting in describing the intermediate states as nucleons in a relativistic Fermi gas, preserving the complexity and consistency in the initial and final states. These approximations simplify the calculations considerably, allowing us to provide outcomes in a reasonable computational time. The results obtained under these approximations are validated by comparing with those from the full model. Additionally, the theoretical predictions are compared with experimental data of the longitudinal and transverse responses of carbon 12. The agreement with data is outstanding for the longitudinal response, where the contribution from the two-body operator is negligible. In the transverse sector, the two-body current increases the response from 30 to 15%, depending on the approximations and kinematics, in general, improving the agreement with data., Comment: 11 pages, 5 figures

This paper presents a new robust loss function, the T-Loss, for medical image segmentation. The proposed loss is based on the negative log-likelihood of the Student-t distribution and can effectively handle outliers in the data by controlling its sensitivity with a single parameter. This parameter is updated during the backpropagation process, eliminating the need for additional computation or prior information about the level and spread of noisy labels. Our experiments show that the T-Loss outperforms traditional loss functions in terms of dice scores on two public medical datasets for skin lesion and lung segmentation. We also demonstrate the ability of T-Loss to handle different types of simulated label noise, resembling human error. Our results provide strong evidence that the T-Loss is a promising alternative for medical image segmentation where high levels of noise or outliers in the dataset are a typical phenomenon in practice. The project website can be found at https://robust-tloss.github.io, Comment: Early accepted to MICCAI 2023

Benchmark datasets in computer vision often contain off-topic images, near duplicates, and label errors, leading to inaccurate estimates of model performance. In this paper, we revisit the task of data cleaning and formalize it as either a ranking problem, which significantly reduces human inspection effort, or a scoring problem, which allows for automated decisions based on score distributions. We find that a specific combination of context-aware self-supervised representation learning and distance-based indicators is effective in finding issues without annotation biases. This methodology, which we call SelfClean, surpasses state-of-the-art performance in detecting off-topic images, near duplicates, and label errors within widely-used image datasets, such as ImageNet-1k, Food-101N, and STL-10, both for synthetic issues and real contamination. We apply the detailed method to multiple image benchmarks, identify up to 16% of issues, and confirm an improvement in evaluation reliability upon cleaning. The official implementation can be found at: https://github.com/Digital-Dermatology/SelfClean., Comment: Accepted at Neural Information Processing Systems (NeurIPS 2024)

Gas source localization (GSL) with an autonomous robot is a problem with many prospective applications, from finding pipe leaks to emergency-response scenarios. In this work, we present a new method to perform GSL in realistic indoor environments, featuring obstacles and turbulent flow. Given the highly complex relationship between the source position and the measurements available to the robot (the single-point gas concentration, and the wind vector) we propose an observation model that derives from contrasting the online, real-time simulation of the gas dispersion from any candidate source localization against a gas concentration map built from sensor readings. To account for a convenient and grounded integration of both into a probabilistic estimation framework, we introduce the concept of probabilistic gas-hit maps, which provide a higher level of abstraction to model the time-dependent nature of gas dispersion. Results from both simulated and real experiments show the capabilities of our current proposal to deal with source localization in complex indoor environments.

In neutrino event generators, for models for neutrino and electron scattering only inclusive cross sections are implemented. When these models are used to describe a semi-inclusive cross section, the event generator attaches the hadron variables based on some assumptions. In this work we compare the nucleon kinematics given by the method used in the GENIE event generator, e.g. in the implementation of the SuSAv2 model, to a fully unfactorized calculation using the relativistic distorted wave impulse approximation (RDWIA). We focus on kinematics relevant to the $e4\nu$ analysis and show that observables obtained with RDWIA differ significantly from those of the approximate method used in GENIE, the latter should be considered unrealistic., Comment: Submitted as part of the proceedings of the 23rd Intl Workshop on Neutrinos from Accelerators (NuFact2022) [C22-07-30]

In this work we present a comparison of semi-inclusive muon and electron neutrino cross sections with $^{40}$Ar target measured by the MicroBooNE Collaboration with the predictions of an unfactorized model based on the relativistic distorted wave impulse approximation (RDWIA) and the SuSAv2-MEC model implemented in the neutrino event generator GENIE. The predictions based on the RDWIA approach, with a realistic description of the initial state and a phenomenological relativistic complex optical potential for the description of final state interactions, better describe the measured cross sections than GENIE-SuSAv2 and RDWIA with a purely real potential.

Pion production on nuclei constitutes a significant part of the total cross section in experiments involving few-GeV neutrinos. Combined analyses of data on deuterium and heavier nuclei points to tensions between the bubble chamber data and the data of the MINER$\nu$A experiment, which are often ascribed to unspecified nuclear effects. To understand the origin of these tensions, a microscopic quantum mechanical framework is needed to compute nuclear matrix elements. We use the local approximation to the relativistic distorted wave impulse approximation (RDWIA) to assess the role of final-state nucleon distortion. To perform this comparison under conditions relevant to neutrino experiments, we compute cross sections for the MINER$\nu$A and T2K charged pion production datasets. The inclusion of nucleon distortion leads to a reduction of the cross section up to 10\%, but to no significant change in shape of the flux-averaged cross sections. Results with and without distortion compare favorably to experimental data, with the exception of the low-$Q^2$ MINER$\nu$A $\pi^+$ data. We point out that hydrogen target data from BEBC is also overpredicted at low-$Q^2$, and that the discrepancy is similar in shape and magnitude to what is found in comparison to MINER$\nu$A data. Including nucleon distortion alone cannot explain the overprediction of low-$Q^2$ cross sections measured by MINER$\nu$A. The similar overprediction of BEBC data on hydrogen means that it is impossible to ascribe this discrepancy solely to a nuclear effect. Axial couplings and their $Q^2$ dependence should ideally be derived from more precise data on hydrogen and deuterium.

We present a complete comparison of semi-inclusive $\nu_\mu$-$^{12}$C cross-section measurements by T2K and MINER$\nu$A collaborations with the predictions from the SuSAv2-MEC model implemented in the neutrino-nucleus event generator GENIE and an unfactorized approach based on the relativistic distorted wave impulse approximation (RDWIA). Results, that include cross sections as function of the final muon and proton kinematics and correlations between both, show that the agreement with data obtained by the RDWIA approach, that accounts for final-state interactions, matches or improves GENIE-SuSAv2 predictions for very forward angles where scaling violations are relevant.