Pablo Márquez-Neila

Sergio Tascon-Morales, Pablo Márquez-Neila, Raphael Sznitman

Visual Question Answering (VQA) models take an image and a natural-language question as input and infer the answer to the question. Recently, VQA systems in medical imaging have gained popularity thanks to potential advantages such as patient engagement and second opinions for clinicians. While most research efforts have been focused on improving architectures and overcoming data-related limitations, answer consistency has been overlooked even though it plays a critical role in establishing trustworthy models. In this work, we propose a novel loss function and corresponding training procedure that allows the inclusion of relations between questions into the training process. Specifically, we consider the case where implications between perception and reasoning questions are known a-priori. To show the benefits of our approach, we evaluate it on the clinically relevant task of Diabetic Macular Edema (DME) staging from fundus imaging. Our experiments show that our method outperforms state-of-the-art baselines, not only by improving model consistency, but also in terms of overall model accuracy. Our code and data are available at https://github.com/sergiotasconmorales/consistency_vqa.

Sergio Tascon-Morales, Pablo Márquez-Neila, Raphael Sznitman

Visual Question Answering (VQA) models aim to answer natural language questions about given images. Due to its ability to ask questions that differ from those used when training the model, medical VQA has received substantial attention in recent years. However, existing medical VQA models typically focus on answering questions that refer to an entire image rather than where the relevant content may be located in the image. Consequently, VQA models are limited in their interpretability power and the possibility to probe the model about specific image regions. This paper proposes a novel approach for medical VQA that addresses this limitation by developing a model that can answer questions about image regions while considering the context necessary to answer the questions. Our experimental results demonstrate the effectiveness of our proposed model, outperforming existing methods on three datasets. Our code and data are available at https://github.com/sergiotasconmorales/locvqa.

Lukas Zbinden, Lars Doorenbos, Theodoros Pissas et al.

Semantic segmentation has made significant progress in recent years thanks to deep neural networks, but the common objective of generating a single segmentation output that accurately matches the image's content may not be suitable for safety-critical domains such as medical diagnostics and autonomous driving. Instead, multiple possible correct segmentation maps may be required to reflect the true distribution of annotation maps. In this context, stochastic semantic segmentation methods must learn to predict conditional distributions of labels given the image, but this is challenging due to the typically multimodal distributions, high-dimensional output spaces, and limited annotation data. To address these challenges, we propose a conditional categorical diffusion model (CCDM) for semantic segmentation based on Denoising Diffusion Probabilistic Models. Our model is conditioned to the input image, enabling it to generate multiple segmentation label maps that account for the aleatoric uncertainty arising from divergent ground truth annotations. Our experimental results show that CCDM achieves state-of-the-art performance on LIDC, a stochastic semantic segmentation dataset, and outperforms established baselines on the classical segmentation dataset Cityscapes.

Lars Doorenbos, Pablo Márquez-Neila, Raphael Sznitman et al.

Hyperbolic space is becoming a popular choice for representing data due to the hierarchical structure - whether implicit or explicit - of many real-world datasets. Along with it comes a need for algorithms capable of solving fundamental tasks, such as classification, in hyperbolic space. Recently, multiple papers have investigated hyperbolic alternatives to hyperplane-based classifiers, such as logistic regression and SVMs. While effective, these approaches struggle with more complex hierarchical data. We, therefore, propose to generalize the well-known random forests to hyperbolic space. We do this by redefining the notion of a split using horospheres. Since finding the globally optimal split is computationally intractable, we find candidate horospheres through a large-margin classifier. To make hyperbolic random forests work on multi-class data and imbalanced experiments, we furthermore outline a new method for combining classes based on their lowest common ancestor and a class-balanced version of the large-margin loss. Experiments on standard and new benchmarks show that our approach outperforms both conventional random forest algorithms and recent hyperbolic classifiers.

Sergio Tascon-Morales, Pablo Márquez-Neila, Raphael Sznitman

Despite considerable recent progress in Visual Question Answering (VQA) models, inconsistent or contradictory answers continue to cast doubt on their true reasoning capabilities. However, most proposed methods use indirect strategies or strong assumptions on pairs of questions and answers to enforce model consistency. Instead, we propose a novel strategy intended to improve model performance by directly reducing logical inconsistencies. To do this, we introduce a new consistency loss term that can be used by a wide range of the VQA models and which relies on knowing the logical relation between pairs of questions and answers. While such information is typically not available in VQA datasets, we propose to infer these logical relations using a dedicated language model and use these in our proposed consistency loss function. We conduct extensive experiments on the VQA Introspect and DME datasets and show that our method brings improvements to state-of-the-art VQA models, while being robust across different architectures and settings.

Alain Jungo, Lars Doorenbos, Tommaso Da Col et al.

Purpose: A fundamental problem in designing safe machine learning systems is identifying when samples presented to a deployed model differ from those observed at training time. Detecting so-called out-of-distribution (OoD) samples is crucial in safety-critical applications such as robotically guided retinal microsurgery, where distances between the instrument and the retina are derived from sequences of 1D images that are acquired by an instrument-integrated optical coherence tomography (iiOCT) probe. Methods: This work investigates the feasibility of using an OoD detector to identify when images from the iiOCT probe are inappropriate for subsequent machine learning-based distance estimation. We show how a simple OoD detector based on the Mahalanobis distance can successfully reject corrupted samples coming from real-world ex vivo porcine eyes. Results: Our results demonstrate that the proposed approach can successfully detect OoD samples and help maintain the performance of the downstream task within reasonable levels. MahaAD outperformed a supervised approach trained on the same kind of corruptions and achieved the best performance in detecting OoD cases from a collection of iiOCT samples with real-world corruptions. Conclusion: The results indicate that detecting corrupted iiOCT data through OoD detection is feasible and does not need prior knowledge of possible corruptions. Consequently, MahaAD could aid in ensuring patient safety during robotically guided microsurgery by preventing deployed prediction models from estimating distances that put the patient at risk.

Sergio Tascon-Morales, Pablo Márquez-Neila, Raphael Sznitman

With growing interest in recent years, medical visual question answering (Med-VQA) has rapidly evolved, with multimodal large language models (MLLMs) emerging as an alternative to classical model architectures. Specifically, their ability to add visual information to the input of pre-trained LLMs brings new capabilities for image interpretation. However, simple visual errors cast doubt on the actual visual understanding abilities of these models. To address this, region-based questions have been proposed as a means to assess and enhance actual visual understanding through compositional evaluation. To combine these two perspectives, this paper introduces targeted visual prompting to equip MLLMs with region-based questioning capabilities. By presenting the model with both the isolated region and the region in its context in a customized visual prompt, we show the effectiveness of our method across multiple datasets while comparing it to several baseline models. Our code and data are available at https://github.com/sergiotasconmorales/locvqallm.

Lars Doorenbos, Olena Torbaniuk, Stefano Cavuoti et al.

Modern sky surveys are producing ever larger amounts of observational data, which makes the application of classical approaches for the classification and analysis of objects challenging and time-consuming. However, this issue may be significantly mitigated by the application of automatic machine and deep learning methods. We propose ULISSE, a new deep learning tool that, starting from a single prototype object, is capable of identifying objects sharing the same morphological and photometric properties, and hence of creating a list of candidate sosia. In this work, we focus on applying our method to the detection of AGN candidates in a Sloan Digital Sky Survey galaxy sample, since the identification and classification of Active Galactic Nuclei (AGN) in the optical band still remains a challenging task in extragalactic astronomy. Intended for the initial exploration of large sky surveys, ULISSE directly uses features extracted from the ImageNet dataset to perform a similarity search. The method is capable of rapidly identifying a list of candidates, starting from only a single image of a given prototype, without the need for any time-consuming neural network training. Our experiments show ULISSE is able to identify AGN candidates based on a combination of host galaxy morphology, color and the presence of a central nuclear source, with a retrieval efficiency ranging from 21% to 65% (including composite sources) depending on the prototype, where the random guess baseline is 12%. We find ULISSE to be most effective in retrieving AGN in early-type host galaxies, as opposed to prototypes with spiral- or late-type properties. Based on the results described in this work, ULISSE can be a promising tool for selecting different types of astrophysical objects in current and future wide-field surveys (e.g. Euclid, LSST etc.) that target millions of sources every single night.

Lars Doorenbos, Raphael Sznitman, Pablo Márquez-Neila

The inability of deep learning models to handle data drawn from unseen distributions has sparked much interest in unsupervised out-of-distribution (U-OOD) detection, as it is crucial for reliable deep learning models. Despite considerable attention, theoretically-motivated approaches are few and far between, with most methods building on top of some form of heuristic. Recently, U-OOD was formalized in the context of data invariants, allowing a clearer understanding of how to characterize U-OOD, and methods leveraging affine invariants have attained state-of-the-art results on large-scale benchmarks. Nevertheless, the restriction to affine invariants hinders the expressiveness of the approach. In this work, we broaden the affine invariants formulation to a more general case and propose a framework consisting of a normalizing flow-like architecture capable of learning non-linear invariants. Our novel approach achieves state-of-the-art results on an extensive U-OOD benchmark, and we demonstrate its further applicability to tabular data. Finally, we show our method has the same desirable properties as those based on affine invariants.

Theodoros Pissas, Pablo Márquez-Neila, Sebastian Wolf et al.

This work explores the effectiveness of masked image modelling for learning representations of retinal OCT images. To this end, we leverage Masked Autoencoders (MAE), a simple and scalable method for self-supervised learning, to obtain a powerful and general representation for OCT images by training on 700K OCT images from 41K patients collected under real world clinical settings. We also provide the first extensive evaluation for a model of OCT on a challenging battery of 6 downstream tasks. Our model achieves strong performance when fully finetuned but can also serve as a versatile frozen feature extractor for many tasks using lightweight adapters. Furthermore, we propose an extension of the MAE pretraining to fuse OCT with an auxiliary modality, namely, IR fundus images and learn a joint model for both. We demonstrate our approach improves performance on a multimodal downstream application. Our experiments utilize most publicly available OCT datasets, thus enabling future comparisons. Our code and model weights are publicly available https://github.com/TheoPis/MIM_OCT.

Lars Doorenbos, Raphael Sznitman, Pablo Márquez-Neila

The reliability of supervised classifiers is severely hampered by their limitations in dealing with unexpected inputs, leading to great interest in out-of-distribution (OOD) detection. Recently, OOD detectors trained on synthetic outliers, especially those generated by large diffusion models, have shown promising results in defining robust OOD decision boundaries. Building on this progress, we present NCIS, which enhances the quality of synthetic outliers by operating directly in the diffusion's model embedding space rather than combining disjoint models as in previous work and by modeling class-conditional manifolds with a conditional volume-preserving network for more expressive characterization of the training distribution. We demonstrate that these improvements yield new state-of-the-art OOD detection results on standard ImageNet100 and CIFAR100 benchmarks and provide insights into the importance of data pre-processing and other key design choices. We make our code available at \url{https://github.com/LarsDoorenbos/NCIS}.

Lars Doorenbos, C. H. Lucas Patty, Raphael Sznitman et al.

Mueller matrices (MMs) encode information on geometry and material properties, but recovering both simultaneously is an ill-posed problem. We explore whether MMs contain sufficient information to infer surface geometry and material properties with machine learning. We use a dataset of spheres of various isotropic materials, with MMs captured over the full angular domain at five visible wavelengths (450-650 nm). We train machine learning models to predict material properties and surface normals using only these MMs as input. We demonstrate that, even when the material type is unknown, surface normals can be predicted and object geometry reconstructed. Moreover, MMs allow models to identify material types correctly. Further analyses show that diagonal elements are key for material characterization, and off-diagonal elements are decisive for normal estimation.

Lars Doorenbos, Eva Sextl, Kevin Heng et al.

Modern spectroscopic surveys can only target a small fraction of the vast amount of photometrically cataloged sources in wide-field surveys. Here, we report the development of a generative AI method capable of predicting optical galaxy spectra from photometric broad-band images alone. This method draws from the latest advances in diffusion models in combination with contrastive networks. We pass multi-band galaxy images into the architecture to obtain optical spectra. From these, robust values for galaxy properties can be derived with any methods in the spectroscopic toolbox, such as standard population synthesis techniques and Lick indices. When trained and tested on 64x64-pixel images from the Sloan Digital Sky Survey, the global bimodality of star-forming and quiescent galaxies in photometric space is recovered, as well as a mass-metallicity relation of star-forming galaxies. The comparison between the observed and the artificially created spectra shows good agreement in overall metallicity, age, Dn4000, stellar velocity dispersion, and E(B-V) values. Photometric redshift estimates of our generative algorithm can compete with other current, specialized deep-learning techniques. Moreover, this work is the first attempt in the literature to infer velocity dispersion from photometric images. Additionally, we can predict the presence of an active galactic nucleus up to an accuracy of 82%. With our method, scientifically interesting galaxy properties, normally requiring spectroscopic inputs, can be obtained in future data sets from large-scale photometric surveys alone. The spectra prediction via AI can further assist in creating realistic mock catalogs.

Lars Doorenbos, Raphael Sznitman, Pablo Márquez-Neila

Deep learning models are vulnerable to performance degradation when encountering out-of-distribution (OOD) images, potentially leading to misdiagnoses and compromised patient care. These shortcomings have led to great interest in the field of OOD detection. Existing unsupervised OOD (U-OOD) detection methods typically assume that OOD samples originate from an unconcentrated distribution complementary to the training distribution, neglecting the reality that deployed models passively accumulate task-specific OOD samples over time. To better reflect this real-world scenario, we introduce Iterative Deployment Exposure (IDE), a novel and more realistic setting for U-OOD detection. We propose CSO, a method for IDE that starts from a U-OOD detector that is agnostic to the OOD distribution and slowly refines it during deployment using observed unlabeled data. CSO uses a new U-OOD scoring function that combines the Mahalanobis distance with a nearest-neighbor approach, along with a novel confidence-scaled few-shot OOD detector to effectively learn from limited OOD examples. We validate our approach on a dedicated benchmark, showing that our method greatly improves upon strong baselines on three medical imaging modalities.

Lars Doorenbos, Raphael Sznitman, Pablo Márquez-Neila

Unsupervised out-of-distribution (U-OOD) detection has recently attracted much attention due its importance in mission-critical systems and broader applicability over its supervised counterpart. Despite this increase in attention, U-OOD methods suffer from important shortcomings. By performing a large-scale evaluation on different benchmarks and image modalities, we show in this work that most popular state-of-the-art methods are unable to consistently outperform a simple anomaly detector based on pre-trained features and the Mahalanobis distance (MahaAD). A key reason for the inconsistencies of these methods is the lack of a formal description of U-OOD. Motivated by a simple thought experiment, we propose a characterization of U-OOD based on the invariants of the training dataset. We show how this characterization is unknowingly embodied in the top-scoring MahaAD method, thereby explaining its quality. Furthermore, our approach can be used to interpret predictions of U-OOD detectors and provides insights into good practices for evaluating future U-OOD methods.

Leonardo Citraro, Pablo Márquez-Neila, Stefano Savarè et al.

Given an image sequence featuring a portion of a sports field filmed by a moving and uncalibrated camera, such as the one of the smartphones, our goal is to compute automatically in real time the focal length and extrinsic camera parameters for each image in the sequence without using a priori knowledges of the position and orientation of the camera. To this end, we propose a novel framework that combines accurate localization and robust identification of specific keypoints in the image by using a fully convolutional deep architecture. Our algorithm exploits both the field lines and the players' image locations, assuming their ground plane positions to be given, to achieve accuracy and robustness that is beyond the current state of the art. We will demonstrate its effectiveness on challenging soccer, basketball, and volleyball benchmark datasets.

Thomas Kurmann, Pablo Márquez-Neila, Siqing Yu et al.

Optical Coherence Tomography (OCT) is the primary imaging modality for detecting pathological biomarkers associated to retinal diseases such as Age-Related Macular Degeneration. In practice, clinical diagnosis and treatment strategies are closely linked to biomarkers visible in OCT volumes and the ability to identify these plays an important role in the development of ophthalmic pharmaceutical products. In this context, we present a method that automatically predicts the presence of biomarkers in OCT cross-sections by incorporating information from the entire volume. We do so by adding a bidirectional LSTM to fuse the outputs of a Convolutional Neural Network that predicts individual biomarkers. We thus avoid the need to use pixel-wise annotations to train our method, and instead provide fine-grained biomarker information regardless. On a dataset of 416 volumes, we show that our approach imposes coherence between biomarker predictions across volume slices and our predictions are superior to several existing approaches.

Andrea Mendizabal, Pablo Márquez-Neila, Stéphane Cotin

The finite element method (FEM) is among the most commonly used numerical methods for solving engineering problems. Due to its computational cost, various ideas have been introduced to reduce computation times, such as domain decomposition, parallel computing, adaptive meshing, and model order reduction. In this paper we present U-Mesh: a data-driven method based on a U-Net architecture that approximates the non-linear relation between a contact force and the displacement field computed by a FEM algorithm. We show that deep learning, one of the latest machine learning methods based on artificial neural networks, can enhance computational mechanics through its ability to encode highly non-linear models in a compact form. Our method is applied to two benchmark examples: a cantilever beam and an L-shape subject to moving punctual loads. A comparison between our method and proper orthogonal decomposition (POD) is done through the paper. The results show that U-Mesh can perform very fast simulations on various geometries, mesh resolutions and number of input forces with very small errors.

Fayez Lahoud, Radhakrishna Achanta, Pablo Márquez-Neila et al.

We present a method to train self-binarizing neural networks, that is, networks that evolve their weights and activations during training to become binary. To obtain similar binary networks, existing methods rely on the sign activation function. This function, however, has no gradients for non-zero values, which makes standard backpropagation impossible. To circumvent the difficulty of training a network relying on the sign activation function, these methods alternate between floating-point and binary representations of the network during training, which is sub-optimal and inefficient. We approach the binarization task by training on a unique representation involving a smooth activation function, which is iteratively sharpened during training until it becomes a binary representation equivalent to the sign activation function. Additionally, we introduce a new technique to perform binary batch normalization that simplifies the conventional batch normalization by transforming it into a simple comparison operation. This is unlike existing methods, which are forced to the retain the conventional floating-point-based batch normalization. Our binary networks, apart from displaying advantages of lower memory and computation as compared to conventional floating-point and binary networks, also show higher classification accuracy than existing state-of-the-art methods on multiple benchmark datasets.

Pablo Márquez-Neila, Mathieu Salzmann, Pascal Fua

Imposing constraints on the output of a Deep Neural Net is one way to improve the quality of its predictions while loosening the requirements for labeled training data. Such constraints are usually imposed as soft constraints by adding new terms to the loss function that is minimized during training. An alternative is to impose them as hard constraints, which has a number of theoretical benefits but has not been explored so far due to the perceived intractability of the problem. In this paper, we show that imposing hard constraints can in fact be done in a computationally feasible way and delivers reasonable results. However, the theoretical benefits do not materialize and the resulting technique is no better than existing ones relying on soft constraints. We analyze the reasons for this and hope to spur other researchers into proposing better solutions.

Radhakrishna Achanta, Pablo Márquez-Neila, Pascal Fua et al.

Size uniformity is one of the main criteria of superpixel methods. But size uniformity rarely conforms to the varying content of an image. The chosen size of the superpixels therefore represents a compromise - how to obtain the fewest superpixels without losing too much important detail. We propose that a more appropriate criterion for creating image segments is information uniformity. We introduce a novel method for segmenting an image based on this criterion. Since information is a natural way of measuring image complexity, our proposed algorithm leads to image segments that are smaller and denser in areas of high complexity and larger in homogeneous regions, thus simplifying the image while preserving its details. Our algorithm is simple and requires just one input parameter - a threshold on the information content. On segmentation comparison benchmarks it proves to be superior to the state-of-the-art. In addition, our method is computationally very efficient, approaching real-time performance, and is easily extensible to three-dimensional image stacks and video volumes.

Bugra Tekin, Pablo Márquez-Neila, Mathieu Salzmann et al.

Most recent approaches to monocular 3D human pose estimation rely on Deep Learning. They typically involve regressing from an image to either 3D joint coordinates directly or 2D joint locations from which 3D coordinates are inferred. Both approaches have their strengths and weaknesses and we therefore propose a novel architecture designed to deliver the best of both worlds by performing both simultaneously and fusing the information along the way. At the heart of our framework is a trainable fusion scheme that learns how to fuse the information optimally instead of being hand-designed. This yields significant improvements upon the state-of-the-art on standard 3D human pose estimation benchmarks.