Matteo Matteucci

Stefano Samele, Eugenio Lomurno, Teodora Jovanovic et al.

Industrial anomaly detection has historically been a unimodal task. Recent multimodal vision-language models have produced systems that admit textual input alongside the image and are presented as enabling text-guided zero- and few-shot inspection. Yet these methods are evaluated with protocols inherited from unimodal benchmarks that hold the textual condition constant and therefore cannot measure whether language conditions the decision; whether reported gains reflect text guidance or strong pretrained visual features remains open. We introduce Text-Guided Anomaly Detection (TGAD), a structured benchmark that progressively increases the functional role of language across three scenarios: a controlled prompt-sensitivity setting on MVTec AD; a component-tagged extension of MVTec AD that requires the model to restrict its assessment to an instructed part; and the new Assembled Panel Dataset (APD), a realistic industrial setting that requires both defect-type and component-location knowledge. We evaluate one representative model per paradigm: generative large vision-language, training-free discriminative, and embedding-adaptive discriminative. In all three, the textual interface conditions the decision only superficially: prompt content is absorbed unless the object noun is removed (the generative model's I-AUROC drops from 97.4 to 82.6); component-level instructions do not constrain the decision once defects outside the instructed part are admitted as normal (from 90.3 to 66.3); and when both combine on APD, image-level discrimination collapses below the MVTec level, in one case below chance (71.2, 50.5, 31.5). These results suggest that standard benchmarks overstate the text-guided capabilities of current multimodal anomaly detection systems, and that a protocol of this kind is a prerequisite for models that can be reliably controlled through language for industrial deployment.

Lorenzo Cazzella, Francesco Linsalata, Maurizio Magarini et al.

Digital Twins (DTs) for physical wireless environments have been recently proposed as accurate virtual representations of the propagation environment that can enable multi-layer decisions at the physical communication equipment. At high-frequency bands, DTs can help to overcome the challenges emerging in high mobility conditions featuring vehicular environments. In this paper, we propose a novel data-driven workflow for the creation of the DT of a Vehicle-to-Everything (V2X) communication scenario and a multi-modal simulation framework for the generation of realistic sensor data and accurate mmWave/sub-THz wireless channels. The proposed method leverages an automotive simulation and testing framework and an accurate ray-tracing channel simulator. Simulations over an urban scenario show the achievable realistic sensor and channel modelling both at the infrastructure and at ego-vehicles. We showcase the proposed framework on the DT-aided blockage handover task for V2X link restoration, leveraging the framework's dynamic channel generation capabilities for realistic vehicular blockage simulation.

Andrea Falanti, Eugenio Lomurno, Danilo Ardagna et al.

The automated machine learning (AutoML) field has become increasingly relevant in recent years. These algorithms can develop models without the need for expert knowledge, facilitating the application of machine learning techniques in the industry. Neural Architecture Search (NAS) exploits deep learning techniques to autonomously produce neural network architectures whose results rival the state-of-the-art models hand-crafted by AI experts. However, this approach requires significant computational resources and hardware investments, making it less appealing for real-usage applications. This article presents the third version of Pareto-Optimal Progressive Neural Architecture Search (POPNASv3), a new sequential model-based optimization NAS algorithm targeting different hardware environments and multiple classification tasks. Our method is able to find competitive architectures within large search spaces, while keeping a flexible structure and data processing pipeline to adapt to different tasks. The algorithm employs Pareto optimality to reduce the number of architectures sampled during the search, drastically improving the time efficiency without loss in accuracy. The experiments performed on images and time series classification datasets provide evidence that POPNASv3 can explore a large set of assorted operators and converge to optimal architectures suited for the type of data provided under different scenarios.

Andrea Falanti, Eugenio Lomurno, Stefano Samele et al.

Automating the research for the best neural network model is a task that has gained more and more relevance in the last few years. In this context, Neural Architecture Search (NAS) represents the most effective technique whose results rival the state of the art hand-crafted architectures. However, this approach requires a lot of computational capabilities as well as research time, which makes prohibitive its usage in many real-world scenarios. With its sequential model-based optimization strategy, Progressive Neural Architecture Search (PNAS) represents a possible step forward to face this resources issue. Despite the quality of the found network architectures, this technique is still limited in research time. A significant step in this direction has been done by Pareto-Optimal Progressive Neural Architecture Search (POPNAS), which expands PNAS with a time predictor to enable a trade-off between search time and accuracy, considering a multi-objective optimization problem. This paper proposes a new version of the Pareto-Optimal Progressive Neural Architecture Search, called POPNASv2. Our approach enhances its first version and improves its performance. We expanded the search space by adding new operators and improved the quality of both predictors to build more accurate Pareto fronts. Moreover, we introduced cell equivalence checks and enriched the search strategy with an adaptive greedy exploration step. Our efforts allow POPNASv2 to achieve PNAS-like performance with an average 4x factor search time speed-up.

Cristiano Battistini, Riccardo Andrea Izzo, Gianluca Bardaro et al.

Large and small language models have been widely used for robotic task planning. At the same time, vision-language models (VLMs) have successfully tackled problems such as image captioning, scene understanding, and visual question answering. In this work, we combine these two approaches by deploying a compact, open-source multimodal model to generate behavior trees for robotic task planning. The main obstacle to achieving this goal is the lack of an existing dataset that links visual observations and instructions to executable behavior trees. We propose a method to construct such a dataset starting from existing robotic episodes (i.e., Open X-Embodiment), in which a large model serves as a teacher in a multi-stage generation pipeline. We use this dataset to fine-tune VLMs ranging from 500M to 4B parameters via parameter-efficient fine-tuning (PEFT). The generated behavior trees, compatible with the BehaviorTree.CPP library, are evaluated both offline, using structural and lexical metrics, and online through the execution of household tasks in a state-of-the-art embodied simulator. Our results demonstrate that our fine-tuned 4B-parameter VLM approaches the performance of state-of-the-art closed-source models, achieving an 87\% success rate while requiring only a fraction of the computational resources.

Alberto Archetti, Francesca Ieva, Matteo Matteucci

Survival analysis is a fundamental tool in medicine, modeling the time until an event of interest occurs in a population. However, in real-world applications, survival data are often incomplete, censored, distributed, and confidential, especially in healthcare settings where privacy is critical. The scarcity of data can severely limit the scalability of survival models to distributed applications that rely on large data pools. Federated learning is a promising technique that enables machine learning models to be trained on multiple datasets without compromising user privacy, making it particularly well-suited for addressing the challenges of survival data and large-scale survival applications. Despite significant developments in federated learning for classification and regression, many directions remain unexplored in the context of survival analysis. In this work, we propose an extension of the Federated Survival Forest algorithm, called FedSurF++. This federated ensemble method constructs random survival forests in heterogeneous federations. Specifically, we investigate several new tree sampling methods from client forests and compare the results with state-of-the-art survival models based on neural networks. The key advantage of FedSurF++ is its ability to achieve comparable performance to existing methods while requiring only a single communication round to complete. The extensive empirical investigation results in a significant improvement from the algorithmic and privacy preservation perspectives, making the original FedSurF algorithm more efficient, robust, and private. We also present results on two real-world datasets demonstrating the success of FedSurF++ in real-world healthcare studies. Our results underscore the potential of FedSurF++ to improve the scalability and effectiveness of survival analysis in distributed settings while preserving user privacy.

Agnese Chiatti, Riccardo Bertoglio, Nico Catalano et al.

Mobile robots will play a crucial role in the transition towards sustainable agriculture. To autonomously and effectively monitor the state of plants, robots ought to be equipped with visual perception capabilities that are robust to the rapid changes that characterise agricultural settings. In this paper, we focus on the challenging task of segmenting grape bunches from images collected by mobile robots in vineyards. In this context, we present the first study that applies surgical fine-tuning to instance segmentation tasks. We show how selectively tuning only specific model layers can support the adaptation of pre-trained Deep Learning models to newly-collected grape images that introduce visual domain shifts, while also substantially reducing the number of tuned parameters.

Davide Tateo, Davide Antonio Cucci, Matteo Matteucci et al.

Efficient object level representation for monocular semantic simultaneous localization and mapping (SLAM) still lacks a widely accepted solution. In this paper, we propose the use of an efficient representation, based on structural points, for the geometry of objects to be used as landmarks in a monocular semantic SLAM system based on the pose-graph formulation. In particular, an inverse depth parametrization is proposed for the landmark nodes in the pose-graph to store object position, orientation and size/scale. The proposed formulation is general and it can be applied to different geometries; in this paper we focus on indoor environments where human-made artifacts commonly share a planar rectangular shape, e.g., windows, doors, cabinets, etc. The approach can be easily extended to urban scenarios where similar shapes exists as well. Experiments in simulation show good performance, particularly in object geometry reconstruction.

Andrea Lampis, Michela Carlotta Massi, Nicola Pirastu et al.

Polygenic risk scores and other genomic analyses require large individual-level genotype datasets, yet strict data access restrictions impede sharing. Synthetic genotype generation offers a privacy-preserving alternative, but most existing methods operate unconditionally, producing samples without phenotype alignment, or rely on unsupervised compression, creating a gap between statistical fidelity and downstream task utility. We present SNPgen, a two-stage conditional latent diffusion framework for generating phenotype-supervised synthetic genotypes. SNPgen combines GWAS-guided variant selection (1,024-2,048 trait-associated SNPs) with a variational autoencoder for genotype compression and a latent diffusion model conditioned on binary disease labels via classifier-free guidance. Evaluated on 458,724 UK Biobank individuals across four complex diseases (coronary artery disease, breast cancer, type 1 and type 2 diabetes), models trained on synthetic data matched real-data predictive performance in a train-on-synthetic, test-on-real protocol, approaching genome-wide PRS methods that use $2$-$6\times$ more variants. Privacy analysis confirmed zero identical matches, near-random membership inference (AUC $\approx 0.50$), preserved linkage disequilibrium structure, and high allele frequency correlation ($r \geq 0.95$) with source data. A controlled simulation with known causal effects verified faithful recovery of the imposed genetic association structure.

Alberto Archetti, Eugenio Lomurno, Francesco Lattari et al.

Survival analysis studies time-modeling techniques for an event of interest occurring for a population. Survival analysis found widespread applications in healthcare, engineering, and social sciences. However, the data needed to train survival models are often distributed, incomplete, censored, and confidential. In this context, federated learning can be exploited to tremendously improve the quality of the models trained on distributed data while preserving user privacy. However, federated survival analysis is still in its early development, and there is no common benchmarking dataset to test federated survival models. This work provides a novel technique for constructing realistic heterogeneous datasets by starting from existing non-federated datasets in a reproducible way. Specifically, we propose two dataset-splitting algorithms based on the Dirichlet distribution to assign each data sample to a carefully chosen client: quantity-skewed splitting and label-skewed splitting. Furthermore, these algorithms allow for obtaining different levels of heterogeneity by changing a single hyperparameter. Finally, numerical experiments provide a quantitative evaluation of the heterogeneity level using log-rank tests and a qualitative analysis of the generated splits. The implementation of the proposed methods is publicly available in favor of reproducibility and to encourage common practices to simulate federated environments for survival analysis.

Mattia Sabella, Alberto Archetti, Pietro Pinoli et al.

Tabular machine learning systems are frequently trained on data affected by non-uniform corruption, including noisy measurements, missing entries, and feature-specific biases. In practice, these defects are often documented only through column-level reliability indicators rather than instance-wise quality annotations, limiting the applicability of many robustness and cleaning techniques. We present QuAIL, a quality-informed training mechanism that incorporates feature reliability priors directly into the learning process. QuAIL augments existing models with a learnable feature-modulation layer whose updates are selectively constrained by a quality-dependent proximal regularizer, thereby inducing controlled adaptation across features of varying trustworthiness. This stabilizes optimization under structured corruption without explicit data repair or sample-level reweighting. Empirical evaluation across 50 classification and regression datasets demonstrates that QuAIL consistently improves average performance over neural baselines under both random and value-dependent corruption, with especially robust behavior in low-data and systematically biased settings. These results suggest that incorporating feature reliability information directly into optimization dynamics is a practical and effective approach for resilient tabular learning.

Alberto Archetti, Matteo Matteucci

Survival analysis is a subfield of statistics concerned with modeling the occurrence time of a particular event of interest for a population. Survival analysis found widespread applications in healthcare, engineering, and social sciences. However, real-world applications involve survival datasets that are distributed, incomplete, censored, and confidential. In this context, federated learning can tremendously improve the performance of survival analysis applications. Federated learning provides a set of privacy-preserving techniques to jointly train machine learning models on multiple datasets without compromising user privacy, leading to a better generalization performance. However, despite the widespread development of federated learning in recent AI research, few studies focus on federated survival analysis. In this work, we present a novel federated algorithm for survival analysis based on one of the most successful survival models, the random survival forest. We call the proposed method Federated Survival Forest (FedSurF). With a single communication round, FedSurF obtains a discriminative power comparable to deep-learning-based federated models trained over hundreds of federated iterations. Moreover, FedSurF retains all the advantages of random forests, namely low computational cost and natural handling of missing values and incomplete datasets. These advantages are especially desirable in real-world federated environments with multiple small datasets stored on devices with low computational capabilities. Numerical experiments compare FedSurF with state-of-the-art survival models in federated networks, showing how FedSurF outperforms deep-learning-based federated algorithms in realistic environments with non-identically distributed data.

Eugenio Lomurno, Matteo matteucci

Nowadays, owners and developers of deep learning models must consider stringent privacy-preservation rules of their training data, usually crowd-sourced and retaining sensitive information. The most widely adopted method to enforce privacy guarantees of a deep learning model nowadays relies on optimization techniques enforcing differential privacy. According to the literature, this approach has proven to be a successful defence against several models' privacy attacks, but its downside is a substantial degradation of the models' performance. In this work, we compare the effectiveness of the differentially-private stochastic gradient descent (DP-SGD) algorithm against standard optimization practices with regularization techniques. We analyze the resulting models' utility, training performance, and the effectiveness of membership inference and model inversion attacks against the learned models. Finally, we discuss differential privacy's flaws and limits and empirically demonstrate the often superior privacy-preserving properties of dropout and l2-regularization.

Eugenio Lomurno, Matteo Matteucci

Federated learning has emerged as a paradigm for collaborative learning, enabling the development of robust models without the need to centralise sensitive data. However, conventional federated learning techniques have privacy and security vulnerabilities due to the exposure of models, parameters or updates, which can be exploited as an attack surface. This paper presents Federated Knowledge Recycling (FedKR), a cross-silo federated learning approach that uses locally generated synthetic data to facilitate collaboration between institutions. FedKR combines advanced data generation techniques with a dynamic aggregation process to provide greater security against privacy attacks than existing methods, significantly reducing the attack surface. Experimental results on generic and medical datasets show that FedKR achieves competitive performance, with an average improvement in accuracy of 4.24% compared to training models from local data, demonstrating particular effectiveness in data scarcity scenarios.

Eugenio Lomurno, Simone Toffoli, Davide Di Febbo et al.

The growing global elderly population is expected to increase the prevalence of frailty, posing significant challenges to healthcare systems. Frailty, a syndrome associated with ageing, is characterised by progressive health decline, increased vulnerability to stressors and increased risk of mortality. It represents a significant burden on public health and reduces the quality of life of those affected. The lack of a universally accepted method to assess frailty and a standardised definition highlights a critical research gap. Given this lack and the importance of early prevention, this study presents an innovative approach using an instrumented ink pen to ecologically assess handwriting for age group classification. Content-free handwriting data from 80 healthy participants in different age groups (20-40, 41-60, 61-70 and 70+) were analysed. Fourteen gesture- and tremor-related indicators were computed from the raw data and used in five classification tasks. These tasks included discriminating between adjacent and non-adjacent age groups using Catboost and Logistic Regression classifiers. Results indicate exceptional classifier performance, with accuracy ranging from 82.5% to 97.5%, precision from 81.8% to 100%, recall from 75% to 100% and ROC-AUC from 92.2% to 100%. Model interpretability, facilitated by SHAP analysis, revealed age-dependent sensitivity of temporal and tremor-related handwriting features. Importantly, this classification method offers potential for early detection of abnormal signs of ageing in uncontrolled settings such as remote home monitoring, thereby addressing the critical issue of frailty detection and contributing to improved care for older adults.

Eugenio Lomurno, Matteo Matteucci

Generative artificial intelligence has transformed the generation of synthetic data, providing innovative solutions to challenges like data scarcity and privacy, which are particularly critical in fields such as medicine. However, the effective use of this synthetic data to train high-performance models remains a significant challenge. This paper addresses this issue by introducing Knowledge Recycling (KR), a pipeline designed to optimise the generation and use of synthetic data for training downstream classifiers. At the heart of this pipeline is Generative Knowledge Distillation (GKD), the proposed technique that significantly improves the quality and usefulness of the information provided to classifiers through a synthetic dataset regeneration and soft labelling mechanism. The KR pipeline has been tested on a variety of datasets, with a focus on six highly heterogeneous medical image datasets, ranging from retinal images to organ scans. The results show a significant reduction in the performance gap between models trained on real and synthetic data, with models based on synthetic data outperforming those trained on real data in some cases. Furthermore, the resulting models show almost complete immunity to Membership Inference Attacks, manifesting privacy properties missing in models trained with conventional techniques.

Simone Sarti, Eugenio Lomurno, Matteo Matteucci

Today, artificial neural networks are the state of the art for solving a variety of complex tasks, especially in image classification. Such architectures consist of a sequence of stacked layers with the aim of extracting useful information and having it processed by a classifier to make accurate predictions. However, intermediate information within such models is often left unused. In other cases, such as in edge computing contexts, these architectures are divided into multiple partitions that are made functional by including early exits, i.e. intermediate classifiers, with the goal of reducing the computational and temporal load without extremely compromising the accuracy of the classifications. In this paper, we present Anticipate, Ensemble and Prune (AEP), a new training technique based on weighted ensembles of early exits, which aims at exploiting the information in the structure of networks to maximise their performance. Through a comprehensive set of experiments, we show how the use of this approach can yield average accuracy improvements of up to 15% over traditional training. In its hybrid-weighted configuration, AEP's internal pruning operation also allows reducing the number of parameters by up to 41%, lowering the number of multiplications and additions by 18% and the latency time to make inference by 16%. By using AEP, it is also possible to learn weights that allow early exits to achieve better accuracy values than those obtained from single-output reference models.

Cristian Sbrolli, Paolo Cudrano, Matteo Frosi et al.

In recent years, Denoising Diffusion Probabilistic Models (DDPMs) have demonstrated exceptional performance in various 2D generative tasks. Following this success, DDPMs have been extended to 3D shape generation, surpassing previous methodologies in this domain. While many of these models are unconditional, some have explored the potential of using guidance from different modalities. In particular, image guidance for 3D generation has been explored through the utilization of CLIP embeddings. However, these embeddings are designed to align images and text, and do not necessarily capture the specific details needed for shape generation. To address this limitation and enhance image-guided 3D DDPMs with augmented 3D understanding, we introduce CISP (Contrastive Image-Shape Pre-training), obtaining a well-structured image-shape joint embedding space. Building upon CISP, we then introduce IC3D, a DDPM that harnesses CISP's guidance for 3D shape generation from single-view images. This generative diffusion model outperforms existing benchmarks in both quality and diversity of generated 3D shapes. Moreover, despite IC3D's generative nature, its generated shapes are preferred by human evaluators over a competitive single-view 3D reconstruction model. These properties contribute to a coherent embedding space, enabling latent interpolation and conditioned generation also from out-of-distribution images. We find IC3D able to generate coherent and diverse completions also when presented with occluded views, rendering it applicable in controlled real-world scenarios.

Matteo Balice, Yanik Kunzi, Chenyangguang Zhang et al.

Recent feed-forward 3D gaussian splatting methods have made dramatic progress on individual aspects of 3D scene reconstruction, but no existing method jointly addresses dynamic content, multi-view input, and unknown camera poses in a single feed-forward pass. Methods that handle dynamics either require accurate camera poses or accept only monocular input; pose-free multi-view methods address only static scenes; and per-scene optimization methods bridge some of these gaps but at minutes-to-hours cost per scene. We introduce NoPo4D, the first feed-forward system that addresses this empty quadrant. Building on a pretrained geometry backbone and recent 4D Gaussian frameworks, NoPo4D introduces a velocity decomposition that splits Gaussian motion into per-pixel image-plane shifts and depth changes, allowing direct supervision from pseudo ground-truth optical flow on the 2D component. This sidesteps both the differentiable rendering that couples prior posed methods to pose accuracy and the 3D motion ground truth that prior pose-free methods require. The system is rounded out by a bidirectional motion encoder for cross-view and cross-frame feature aggregation, and view-dependent opacity that mitigates cross-view and cross-timestep Gaussian misalignments. On four multi-view dynamic benchmarks, NoPo4D consistently outperforms prior feed-forward baselines, and with an optional post-optimization stage surpasses per-scene optimization methods, while running orders of magnitude faster.

Alberto Archetti, Eugenio Lomurno, Diego Piccinotti et al.

Survival analysis is a statistical framework for modeling time-to-event data. It plays a pivotal role in medicine, reliability engineering, and social science research, where understanding event dynamics even with few data samples is critical. Recent advancements in machine learning, particularly those employing neural networks and decision trees, have introduced sophisticated algorithms for survival modeling. However, many of these methods rely on restrictive assumptions about the underlying event-time distribution, such as proportional hazard, time discretization, or accelerated failure time. In this study, we propose FPBoost, a survival model that combines a weighted sum of fully parametric hazard functions with gradient boosting. Distribution parameters are estimated with decision trees trained by maximizing the full survival likelihood. We show how FPBoost is a universal approximator of hazard functions, offering full event-time modeling flexibility while maintaining interpretability through the use of well-established parametric distributions. We evaluate concordance and calibration of FPBoost across multiple benchmark datasets, showcasing its robustness and versatility as a new tool for survival estimation.

Mirko Usuelli, Matteo Frosi, Paolo Cudrano et al.

Loop Closure Detection (LCD) is an essential task in robotics and computer vision, serving as a fundamental component for various applications across diverse domains. These applications encompass object recognition, image retrieval, and video analysis. LCD consists in identifying whether a robot has returned to a previously visited location, referred to as a loop, and then estimating the related roto-translation with respect to the analyzed location. Despite the numerous advantages of radar sensors, such as their ability to operate under diverse weather conditions and provide a wider range of view compared to other commonly used sensors (e.g., cameras or LiDARs), integrating radar data remains an arduous task due to intrinsic noise and distortion. To address this challenge, this research introduces RadarLCD, a novel supervised deep learning pipeline specifically designed for Loop Closure Detection using the FMCW Radar (Frequency Modulated Continuous Wave) sensor. RadarLCD, a learning-based LCD methodology explicitly designed for radar systems, makes a significant contribution by leveraging the pre-trained HERO (Hybrid Estimation Radar Odometry) model. Being originally developed for radar odometry, HERO's features are used to select key points crucial for LCD tasks. The methodology undergoes evaluation across a variety of FMCW Radar dataset scenes, and it is compared to state-of-the-art systems such as Scan Context for Place Recognition and ICP for Loop Closure. The results demonstrate that RadarLCD surpasses the alternatives in multiple aspects of Loop Closure Detection.

Eugenio Lomurno, Alberto Archetti, Francesca Ausonio et al.

The remarkable proliferation of deep learning across various industries has underscored the importance of data privacy and security in AI pipelines. As the evolution of sophisticated Membership Inference Attacks (MIAs) threatens the secrecy of individual-specific information used for training deep learning models, Differential Privacy (DP) raises as one of the most utilized techniques to protect models against malicious attacks. However, despite its proven theoretical properties, DP can significantly hamper model performance and increase training time, turning its use impractical in real-world scenarios. Tackling this issue, we present Discriminative Adversarial Privacy (DAP), a novel learning technique designed to address the limitations of DP by achieving a balance between model performance, speed, and privacy. DAP relies on adversarial training based on a novel loss function able to minimise the prediction error while maximising the MIA's error. In addition, we introduce a novel metric named Accuracy Over Privacy (AOP) to capture the performance-privacy trade-off. Finally, to validate our claims, we compare DAP with diverse DP scenarios, providing an analysis of the results from performance, time, and privacy preservation perspectives.

Simone Sarti, Eugenio Lomurno, Andrea Falanti et al.

The use of Neural Architecture Search (NAS) techniques to automate the design of neural networks has become increasingly popular in recent years. The proliferation of devices with different hardware characteristics using such neural networks, as well as the need to reduce the power consumption for their search, has led to the realisation of Once-For-All (OFA), an eco-friendly algorithm characterised by the ability to generate easily adaptable models through a single learning process. In order to improve this paradigm and develop high-performance yet eco-friendly NAS techniques, this paper presents OFAv2, the extension of OFA aimed at improving its performance while maintaining the same ecological advantage. The algorithm is improved from an architectural point of view by including early exits, parallel blocks and dense skip connections. The training process is extended by two new phases called Elastic Level and Elastic Height. A new Knowledge Distillation technique is presented to handle multi-output networks, and finally a new strategy for dynamic teacher network selection is proposed. These modifications allow OFAv2 to improve its accuracy performance on the Tiny ImageNet dataset by up to 12.07% compared to the original version of OFA, while maintaining the algorithm flexibility and advantages.

Eugenio Lomurno, Samuele Mariani, Matteo Monti et al.

Neural Architecture Search (NAS) automates neural network design, reducing dependence on human expertise. While NAS methods are computationally intensive and dataset-specific, auxiliary predictors reduce the models needing training, decreasing search time. This strategy is used to generate architectures satisfying multiple computational constraints. Recently, Transferable NAS has emerged, generalizing the search process from dataset-dependent to task-dependent. In this field, DiffusionNAG is a state-of-the-art method. This diffusion-based approach streamlines computation, generating architectures optimized for accuracy on unseen datasets without further adaptation. However, by focusing solely on accuracy, DiffusionNAG overlooks other crucial objectives like model complexity, computational efficiency, and inference latency -- factors essential for deploying models in resource-constrained environments. This paper introduces the Pareto-Optimal Many-Objective Neural Architecture Generator (POMONAG), extending DiffusionNAG via a many-objective diffusion process. POMONAG simultaneously considers accuracy, number of parameters, multiply-accumulate operations (MACs), and inference latency. It integrates Performance Predictor models to estimate these metrics and guide diffusion gradients. POMONAG's optimization is enhanced by expanding its training Meta-Dataset, applying Pareto Front Filtering, and refining embeddings for conditional generation. These enhancements enable POMONAG to generate Pareto-optimal architectures that outperform the previous state-of-the-art in performance and efficiency. Results were validated on two search spaces -- NASBench201 and MobileNetV3 -- and evaluated across 15 image classification datasets.

Filippo Ghilotti, Samuel Brucker, Nahku Saidy et al.

Unlabeled LiDAR logs, in autonomous driving applications, are inherently a gold mine of dense 3D geometry hiding in plain sight - yet they are almost useless without human labels, highlighting a dominant cost barrier for autonomous-perception research. In this work we tackle this bottleneck by leveraging temporal-geometric consistency across LiDAR sweeps to lift and fuse cues from text and 2D vision foundation models directly into 3D, without any manual input. We introduce an unsupervised multi-modal pseudo-labeling method relying on strong geometric priors learned from temporally accumulated LiDAR maps, alongside with a novel iterative update rule that enforces joint geometric-semantic consistency, and vice-versa detecting moving objects from inconsistencies. Our method simultaneously produces 3D semantic labels, 3D bounding boxes, and dense LiDAR scans, demonstrating robust generalization across three datasets. We experimentally validate that our method compares favorably to existing semantic segmentation and object detection pseudo-labeling methods, which often require additional manual supervision. We confirm that even a small fraction of our geometrically consistent, densified LiDAR improves depth prediction by 51.5% and 22.0% MAE in the 80-150 and 150-250 meters range, respectively.

Riccardo Ughi, Eugenio Lomurno, Matteo Matteucci

The Transformer is a highly successful deep learning model that has revolutionised the world of artificial neural networks, first in natural language processing and later in computer vision. This model is based on the attention mechanism and is able to capture complex semantic relationships between a variety of patterns present in the input data. Precisely because of these characteristics, the Transformer has recently been exploited for time series forecasting problems, assuming a natural adaptability to the domain of continuous numerical series. Despite the acclaimed results in the literature, some works have raised doubts about the robustness and effectiveness of this approach. In this paper, we further investigate the effectiveness of Transformer-based models applied to the domain of time series forecasting, demonstrate their limitations, and propose a set of alternative models that are better performing and significantly less complex. In particular, we empirically show how simplifying Transformer-based forecasting models almost always leads to an improvement, reaching state of the art performance. We also propose shallow models without the attention mechanism, which compete with the overall state of the art in long time series forecasting, and demonstrate their ability to accurately predict time series over extremely long windows. From a methodological perspective, we show how it is always necessary to use a simple baseline to verify the effectiveness of proposed models, and finally, we conclude the paper with a reflection on recent research paths and the opportunity to follow trends and hypes even where it may not be necessary.

Cristian Sbrolli, Matteo Matteucci, Toshihiko Yamasaki

Modern Vision-Language Models (VLMs) exhibit a critical flaw in compositional reasoning, often confusing "a red cube and a blue sphere" with "a blue cube and a red sphere". Disentangling the visual and linguistic roots of these failures is a fundamental challenge for robust evaluation. To enable fine-grained, controllable analysis, we introduce Auto-Comp, a fully automated and synthetic pipeline for generating scalable benchmarks. Its controllable nature is key to dissecting and isolating different reasoning skills. Auto-Comp generates paired images from Minimal (e.g., "a monitor to the left of a bicycle on a white background") and LLM-generated Contextual captions (e.g., "In a brightly lit photography studio, a monitor is positioned to the left of a bicycle"), allowing a controlled A/B test to disentangle core binding ability from visio-linguistic complexity. Our evaluation of 20 VLMs on novel benchmarks for color binding and spatial relations reveals universal compositional failures in both CLIP and SigLIP model families. Crucially, our novel "Confusion Benchmark" reveals a deeper flaw beyond simple attribute swaps: models are highly susceptible to low-entropy distractors (e.g., repeated objects or colors), demonstrating their compositional failures extend beyond known bag-of-words limitations. we uncover a surprising trade-off: visio-linguistic context, which provides global scene cues, aids spatial reasoning but simultaneously hinders local attribute binding by introducing visual clutter. We release the Auto-Comp pipeline to facilitate future benchmark creation, alongside all our generated benchmarks (https://huggingface.co/AutoComp).

Paolo Riva, Leonardo Gargani, Matteo Frosi et al.

As the world of agentic artificial intelligence applied to robotics evolves, the need for agents capable of building and retrieving memories and observations efficiently is increasing. Robots operating in complex environments must build memory structures to enable useful human-robot interactions by leveraging the mnemonic representation of the current operating context. People interacting with robots may expect the embodied agent to provide information about locations, events, or objects, which requires the agent to provide precise answers within human-like inference times to be perceived as responsive. We propose the Embodied Light Graph Retrieval Agent (EmbodiedLGR-Agent), a visual-language model (VLM)-driven agent architecture that constructs dense and efficient representations of robot operating environments. EmbodiedLGR-Agent directly addresses the need for an efficient memory representation of the environment by providing a hybrid building-retrieval approach built on parameter-efficient VLMs that store low-level information about objects and their positions in a semantic graph, while retaining high-level descriptions of the observed scenes with a traditional retrieval-augmented architecture. EmbodiedLGR-Agent is evaluated on the popular NaVQA dataset, achieving state-of-the-art performance in inference and querying times for embodied agents, while retaining competitive accuracy on the global task relative to the current state-of-the-art approaches. Moreover, EmbodiedLGR-Agent was successfully deployed on a physical robot, showing practical utility in real-world contexts through human-robot interaction, while running the visual-language model and the building-retrieval pipeline locally.

Yiran Guo, Simone Mentasti, Xiaofeng Jin et al.

3D semantic occupancy prediction is a cornerstone for embodied AI, enabling agents to perceive dense scene geometry and semantics incrementally from monocular video streams. However, current online frameworks face two critical bottlenecks: the inherent depth ambiguity of monocular estimation that causes "feature bleeding" at object boundaries , and the "cold start" instability where uninitialized temporal fusion layers distort high-quality spatial priors during early training stages. In this paper, we propose SGR-OCC (Soft-Gating and Ray-refinement Occupancy), a unified framework driven by the philosophy of "Inheritance and Evolution". To perfectly inherit monocular spatial expertise, we introduce a Soft-Gating Feature Lifter that explicitly models depth uncertainty via a Gaussian gate to probabilistically suppress background noise. Furthermore, a Dynamic Ray-Constrained Anchor Refinement module simplifies complex 3D displacement searches into efficient 1D depth corrections along camera rays, ensuring sub-voxel adherence to physical surfaces. To ensure stable evolution toward temporal consistency, we employ a Two-Phase Progressive Training Strategy equipped with identity-initialized fusion, effectively resolving the cold start problem and shielding spatial priors from noisy early gradients. Extensive experiments on the EmbodiedOcc-ScanNet and Occ-ScanNet benchmarks demonstrate that SGR-OCC achieves state-of-the-art performance. In local prediction tasks, SGR-OCC achieves a completion IoU of 58.55$\%$ and a semantic mIoU of 49.89$\%$, surpassing the previous best method, EmbodiedOcc++, by 3.65$\%$ and 3.69$\%$ respectively. In challenging embodied prediction tasks, our model reaches 55.72$\%$ SC-IoU and 46.22$\%$ mIoU. Qualitative results further confirm our model's superior capability in preserving structural integrity and boundary sharpness in complex indoor environments.

Marina Mastroleo, Alberto Archetti, Federico Mastroleo et al.

Accurate prediction of time-to-event outcomes is critical for clinical decision-making, treatment planning, and resource allocation in modern healthcare. While classical survival models such as Cox remain widely adopted in standard practice, they rely on restrictive assumptions, including linear covariate relationships and proportional hazards over time, that often fail to capture real-world clinical dynamics. Recent deep learning approaches like DeepSurv and DeepHit offer improved expressivity but sacrifice interpretability, limiting clinical adoption where trust and transparency are paramount. Hybrid models incorporating Kolmogorov-Arnold Networks (KANs), such as CoxKAN, have begun to address this trade-off but remain constrained by the semi-parametric Cox framework. In this work we introduce SurvKAN, a fully parametric, time-continuous survival model based on KAN architectures that eliminates the proportional hazards constraint. SurvKAN treats time as an explicit input to a KAN that directly predicts the log-hazard function, enabling end-to-end training on the full survival likelihood. Our architecture preserves interpretability through learnable univariate functions that indicate how individual features influence risk over time. Extensive experiments on standard survival benchmarks demonstrate that SurvKAN achieves competitive or superior performance compared to classical and state-of-the-art baselines across concordance and calibration metrics. Additionally, interpretability analyses reveal clinically meaningful patterns that align with medical domain knowledge.

Andrea Aspesi, Andrea Simpsi, Aaron Tognoli et al.

Event-based cameras are becoming a popular solution for efficient, low-power eye tracking. Due to the sparse and asynchronous nature of event data, they require less processing power and offer latencies in the microsecond range. However, many existing solutions are limited to validation on powerful GPUs, with no deployment on real embedded devices. In this paper, we present EETnet, a convolutional neural network designed for eye tracking using purely event-based data, capable of running on microcontrollers with limited resources. Additionally, we outline a methodology to train, evaluate, and quantize the network using a public dataset. Finally, we propose two versions of the architecture: a classification model that detects the pupil on a grid superimposed on the original image, and a regression model that operates at the pixel level.

Nico Catalano, Matteo Matteucci

Semantic segmentation, vital for applications ranging from autonomous driving to robotics, faces significant challenges in domains where collecting large annotated datasets is difficult or prohibitively expensive. In such contexts, such as medicine and agriculture, the scarcity of training images hampers progress. Introducing Few-Shot Semantic Segmentation, a novel task in computer vision, which aims at designing models capable of segmenting new semantic classes with only a few examples. This paper consists of a comprehensive survey of Few-Shot Semantic Segmentation, tracing its evolution and exploring various model designs, from the more popular conditional and prototypical networks to the more niche latent space optimization methods, presenting also the new opportunities offered by recent foundational models. Through a chronological narrative, we dissect influential trends and methodologies, providing insights into their strengths and limitations. A temporal timeline offers a visual roadmap, marking key milestones in the field's progression. Complemented by quantitative analyses on benchmark datasets and qualitative showcases of seminal works, this survey equips readers with a deep understanding of the topic. By elucidating current challenges, state-of-the-art models, and prospects, we aid researchers and practitioners in navigating the intricacies of Few-Shot Semantic Segmentation and provide ground for future development.

Paolo Cudrano, Matteo Bellusci, Giuseppe Macino et al.

Accurate road surface classification is crucial for autonomous vehicles (AVs) to optimize driving conditions, enhance safety, and enable advanced road mapping. However, deep learning models for road surface classification suffer from poor generalization when tested on unseen datasets. To update these models with new information, also the original training dataset must be taken into account, in order to avoid catastrophic forgetting. This is, however, inefficient if not impossible, e.g., when the data is collected in streams or large amounts. To overcome this limitation and enable fast and efficient cross-dataset adaptation, we propose to employ continual learning finetuning methods designed to retain past knowledge while adapting to new data, thus effectively avoiding forgetting. Experimental results demonstrate the superiority of this approach over naive finetuning, achieving performance close to fresh retraining. While solving this known problem, we also provide a general description of how the same technique can be adopted in other AV scenarios. We highlight the potential computational and economic benefits that a continual-based adaptation can bring to the AV industry, while also reducing greenhouse emissions due to unnecessary joint retraining.

Mirko Usuelli, David Rapado-Rincon, Gert Kootstra et al.

Autonomous robots in orchards require real-time 3D scene understanding despite repetitive row geometry, seasonal appearance changes, and wind-driven foliage motion. We present AgriGS-SLAM, a Visual--LiDAR SLAM framework that couples direct LiDAR odometry and loop closures with multi-camera 3D Gaussian Splatting (3DGS) rendering. Batch rasterization across complementary viewpoints recovers orchard structure under occlusions, while a unified gradient-driven map lifecycle executed between keyframes preserves fine details and bounds memory. Pose refinement is guided by a probabilistic LiDAR-based depth consistency term, back-propagated through the camera projection to tighten geometry-appearance coupling. We deploy the system on a field platform in apple and pear orchards across dormancy, flowering, and harvesting, using a standardized trajectory protocol that evaluates both training-view and novel-view synthesis to reduce 3DGS overfitting in evaluation. Across seasons and sites, AgriGS-SLAM delivers sharper, more stable reconstructions and steadier trajectories than recent state-of-the-art 3DGS-SLAM baselines while maintaining real-time performance on-tractor. While demonstrated in orchard monitoring, the approach can be applied to other outdoor domains requiring robust multimodal perception.

Michele Carlo La Greca, Mirko Usuelli, Matteo Matteucci

Agriculture, fundamental for human sustenance, faces unprecedented challenges. The need for efficient, human-cooperative, and sustainable farming methods has never been greater. The core contributions of this work involve leveraging Active Vision (AV) techniques and Zero-Shot Learning (ZSL) to improve the robot's ability to perceive and interact with agricultural environment in the context of fruit harvesting. The AV Pipeline implemented within ROS 2 integrates the Next-Best View (NBV) Planning for 3D environment reconstruction through a dynamic 3D Occupancy Map. Our system allows the robotics arm to dynamically plan and move to the most informative viewpoints and explore the environment, updating the 3D reconstruction using semantic information produced through ZSL models. Simulation and real-world experimental results demonstrate our system's effectiveness in complex visibility conditions, outperforming traditional and static predefined planning methods. ZSL segmentation models employed, such as YOLO World + EfficientViT SAM, exhibit high-speed performance and accurate segmentation, allowing flexibility when dealing with semantic information in unknown agricultural contexts without requiring any fine-tuning process.

Nicolo Resmini, Eugenio Lomurno, Cristian Sbrolli et al.

Generative diffusion models have emerged as powerful tools to synthetically produce training data, offering potential solutions to data scarcity and reducing labelling costs for downstream supervised deep learning applications. However, effectively leveraging text-conditioned image generation for building classifier training sets requires addressing key issues: constructing informative textual prompts, adapting generative models to specific domains, and ensuring robust performance. This paper proposes the Text-Conditioned Knowledge Recycling (TCKR) pipeline to tackle these challenges. TCKR combines dynamic image captioning, parameter-efficient diffusion model fine-tuning, and Generative Knowledge Distillation techniques to create synthetic datasets tailored for image classification. The pipeline is rigorously evaluated on ten diverse image classification benchmarks. The results demonstrate that models trained solely on TCKR-generated data achieve classification accuracies on par with (and in several cases exceeding) models trained on real images. Furthermore, the evaluation reveals that these synthetic-data-trained models exhibit substantially enhanced privacy characteristics: their vulnerability to Membership Inference Attacks is significantly reduced, with the membership inference AUC lowered by 5.49 points on average compared to using real training data, demonstrating a substantial improvement in the performance-privacy trade-off. These findings indicate that high-fidelity synthetic data can effectively replace real data for training classifiers, yielding strong performance whilst simultaneously providing improved privacy protection as a valuable emergent property. The code and trained models are available in the accompanying open-source repository.

Francesco Visin, Marco Ciccone, Adriana Romero et al.

We propose a structured prediction architecture, which exploits the local generic features extracted by Convolutional Neural Networks and the capacity of Recurrent Neural Networks (RNN) to retrieve distant dependencies. The proposed architecture, called ReSeg, is based on the recently introduced ReNet model for image classification. We modify and extend it to perform the more challenging task of semantic segmentation. Each ReNet layer is composed of four RNN that sweep the image horizontally and vertically in both directions, encoding patches or activations, and providing relevant global information. Moreover, ReNet layers are stacked on top of pre-trained convolutional layers, benefiting from generic local features. Upsampling layers follow ReNet layers to recover the original image resolution in the final predictions. The proposed ReSeg architecture is efficient, flexible and suitable for a variety of semantic segmentation tasks. We evaluate ReSeg on several widely-used semantic segmentation datasets: Weizmann Horse, Oxford Flower, and CamVid; achieving state-of-the-art performance. Results show that ReSeg can act as a suitable architecture for semantic segmentation tasks, and may have further applications in other structured prediction problems. The source code and model hyperparameters are available on https://github.com/fvisin/reseg.

Eugenio Lomurno, Filippo Balzarini, Francesco Benelle et al.

Diffusion-based generators set the current state of the art for synthetic tabular data. These methods approach but rarely exceed real-data utility, and closing this synthetic-real gap has so far been pursued exclusively at training time, via architectural advances, scaling, and retraining of monolithic generators. The inference-time alternative, i.e., refining the outputs of a pre-trained backbone with parameters left untouched, has remained largely unexplored for tabular synthesis. We introduce TARDIS (Tabular generation through Refinement, Distillation, and Inference-time Sampling), an inference-time refinement framework that operates on a frozen pre-trained backbone, configured per dataset by a Tree-structured Parzen Estimator search over score-level guidance during reverse diffusion, with each trial's objective set by an inner grid search over post-hoc sample selectors and an optional soft-label distillation step. The search space encodes a single mathematical pattern we name Bidirectional Chamfer Refinement (BCR): the symmetric Chamfer functional between synthetic and real samples is minimized both continuously, via a score-level gradient, and discretely, via batch-ranking post-generation. The per-dataset search recovers BCR-aligned configurations on most datasets, evidence for BCR as the dominant refinement pattern. Across 15 binary, multiclass, and regression benchmarks TARDIS achieves a median +8.6% downstream-task improvement over models trained on real data (95% CI [+3.3, +16.4], Wilcoxon p=0.016, 11/15 strict wins) and improves over the TabDiff backbone on all 15 datasets (mean +12.9%, p<10^-4), matching the backbone on manifold fidelity, diversity, and sample-level privacy. Inference-time refinement of a pre-trained tabular diffusion backbone reaches and exceeds real-data utility in 1 to 80 minutes on a single consumer-grade GPU.

Eugenio Lomurno, Matteo D'Oria, Matteo Matteucci

Recent advances in generative artificial intelligence have enabled the creation of high-quality synthetic data that closely mimics real-world data. This paper explores the adaptation of the Stable Diffusion 2.0 model for generating synthetic datasets, using Transfer Learning, Fine-Tuning and generation parameter optimisation techniques to improve the utility of the dataset for downstream classification tasks. We present a class-conditional version of the model that exploits a Class-Encoder and optimisation of key generation parameters. Our methodology led to synthetic datasets that, in a third of cases, produced models that outperformed those trained on real datasets.

Giuseppe Chiari, Davide Galli, Francesco Lattari et al.

Side-channel attacks allow extracting secret information from the execution of cryptographic primitives by correlating the partially known computed data and the measured side-channel signal. However, to set up a successful side-channel attack, the attacker has to perform i) the challenging task of locating the time instant in which the target cryptographic primitive is executed inside a side-channel trace and then ii)the time-alignment of the measured data on that time instant. This paper presents a novel deep-learning technique to locate the time instant in which the target computed cryptographic operations are executed in the side-channel trace. In contrast to state-of-the-art solutions, the proposed methodology works even in the presence of trace deformations obtained through random delay insertion techniques. We validated our proposal through a successful attack against a variety of unprotected and protected cryptographic primitives that have been executed on an FPGA-implemented system-on-chip featuring a RISC-V CPU.

Lorenzo Cazzella, Marouan Mizmizi, Dario Tagliaferri et al.

In Integrated Sensing and Communication (ISAC) systems, matching the radar targets with communication user equipments (UEs) is functional to several communication tasks, such as proactive handover and beam prediction. In this paper, we consider a radar-assisted communication system where a base station (BS) is equipped with a multiple-input-multiple-output (MIMO) radar that has a double aim: (i) associate vehicular radar targets to vehicular equipments (VEs) in the communication beamspace and (ii) predict the beamforming vector for each VE from radar data. The proposed target-to-user (T2U) association consists of two stages. First, vehicular radar targets are detected from range-angle images, and, for each, a beamforming vector is estimated. Then, the inferred per-target beamforming vectors are matched with the ones utilized at the BS for communication to perform target-to-user (T2U) association. Joint multi-target detection and beam inference is obtained by modifying the you only look once (YOLO) model, which is trained over simulated range-angle radar images. Simulation results over different urban vehicular mobility scenarios show that the proposed T2U method provides a probability of correct association that increases with the size of the BS antenna array, highlighting the respective increase of the separability of the VEs in the beamspace. Moreover, we show that the modified YOLO architecture can effectively perform both beam prediction and radar target detection, with similar performance in mean average precision on the latter over different antenna array sizes.

Agnese Chiatti, Sara Bernardini, Lara Shibelski Godoy Piccolo et al.

The rapid adoption of Vision Language Models (VLMs), pre-trained on large image-text and video-text datasets, calls for protecting and informing users about when to trust these systems. This survey reviews studies on trust dynamics in user-VLM interactions, through a multi-disciplinary taxonomy encompassing different cognitive science capabilities, collaboration modes, and agent behaviours. Literature insights and findings from a workshop with prospective VLM users inform preliminary requirements for future VLM trust studies.

Lorenzo Cazzella, Francesco Linsalata, Mahdi Maleki et al.

Wireless communication systems can significantly benefit from the availability of spatially consistent representations of the wireless channel to efficiently perform a wide range of communication tasks. Towards this purpose, channel charting has been introduced as an effective unsupervised learning technique to achieve both locally and globally consistent radio maps. In this letter, we propose Chartwin, a case study on the integration of localization-oriented channel charting with dynamic Digital Network Twins (DNTs). Numerical results showcase the significant performance of semi-supervised channel charting in constructing a spatially consistent chart of the considered extended urban environment. The considered method results in $\approx$ 4.5 m localization error for the static DNT and $\approx$ 6 m in the dynamic DNT, fostering DNT-aided channel charting and localization.

Andrea Menta, Alberto Archetti, Matteo Matteucci

Neural cellular automata represent an evolution of the traditional cellular automata model, enhanced by the integration of a deep learning-based transition function. This shift from a manual to a data-driven approach significantly increases the adaptability of these models, enabling their application in diverse domains, including content generation and artificial life. However, their widespread application has been hampered by significant computational requirements. In this work, we introduce the Latent Neural Cellular Automata (LNCA) model, a novel architecture designed to address the resource limitations of neural cellular automata. Our approach shifts the computation from the conventional input space to a specially designed latent space, relying on a pre-trained autoencoder. We apply our model in the context of image restoration, which aims to reconstruct high-quality images from their degraded versions. This modification not only reduces the model's resource consumption but also maintains a flexible framework suitable for various applications. Our model achieves a significant reduction in computational requirements while maintaining high reconstruction fidelity. This increase in efficiency allows for inputs up to 16 times larger than current state-of-the-art neural cellular automata models, using the same resources.

Agnese Taluzzi, Davide Gesualdi, Riccardo Santambrogio et al.

This report presents SceneNet and KnowledgeNet, our approaches developed for the HD-EPIC VQA Challenge 2025. SceneNet leverages scene graphs generated with a multi-modal large language model (MLLM) to capture fine-grained object interactions, spatial relationships, and temporally grounded events. In parallel, KnowledgeNet incorporates ConceptNet's external commonsense knowledge to introduce high-level semantic connections between entities, enabling reasoning beyond directly observable visual evidence. Each method demonstrates distinct strengths across the seven categories of the HD-EPIC benchmark, and their combination within our framework results in an overall accuracy of 44.21% on the challenge, highlighting its effectiveness for complex egocentric VQA tasks.

Xiaofeng Jin, Matteo Frosi, Matteo Matteucci

Real-time open-vocabulary scene understanding is essential for efficient 3D perception in applications such as vision-language navigation, embodied intelligence, and augmented reality. However, existing methods suffer from imprecise instance segmentation, static semantic updates, and limited handling of complex queries. To address these issues, we present OpenFusion++, a TSDF-based real-time 3D semantic-geometric reconstruction system. Our approach refines 3D point clouds by fusing confidence maps from foundational models, dynamically updates global semantic labels via an adaptive cache based on instance area, and employs a dual-path encoding framework that integrates object attributes with environmental context for precise query responses. Experiments on the ICL, Replica, ScanNet, and ScanNet++ datasets demonstrate that OpenFusion++ significantly outperforms the baseline in both semantic accuracy and query responsiveness.

Xiaofeng Jin, Yan Fang, Matteo Frosi et al.

Scene view synthesis, which generates novel views from limited perspectives, is increasingly vital for applications like virtual reality, augmented reality, and robotics. Unlike object-based tasks, such as generating 360° views of a car, scene view synthesis handles entire environments where non-uniform observations pose unique challenges for stable rendering quality. To address this issue, we propose a novel approach: renderability field-guided gaussian splatting (RF-GS). This method quantifies input inhomogeneity through a renderability field, guiding pseudo-view sampling to enhanced visual consistency. To ensure the quality of wide-baseline pseudo-views, we train an image restoration model to map point projections to visible-light styles. Additionally, our validated hybrid data optimization strategy effectively fuses information of pseudo-view angles and source view textures. Comparative experiments on simulated and real-world data show that our method outperforms existing approaches in rendering stability.

Andrea Simpsi, Andrea Aspesi, Simone Mentasti et al.

Event-based eye tracking is a promising solution for efficient and low-power eye tracking in smart eyewear technologies. However, the novelty of event-based sensors has resulted in a limited number of available datasets, particularly those with eye-level annotations, crucial for algorithm validation and deep-learning training. This paper addresses this gap by presenting an improved version of a popular event-based eye-tracking dataset. We introduce a semi-automatic annotation pipeline specifically designed for event-based data annotation. Additionally, we provide the scientific community with the computed annotations for pupil detection at 200Hz.

Fatemeh Baghdadi, Davide Cirillo, Daniele Lezzi et al.

The AI-SPRINT project, launched in 2021 and funded by the European Commission, focuses on the development and implementation of AI applications across the computing continuum. This continuum ensures the coherent integration of computational resources and services from centralized data centers to edge devices, facilitating efficient and adaptive computation and application delivery. AI-SPRINT has achieved significant scientific advances, including streamlined processes, improved efficiency, and the ability to operate in real time, as evidenced by three practical use cases. This paper provides an in-depth examination of these applications -- Personalized Healthcare, Maintenance and Inspection, and Farming 4.0 -- highlighting their practical implementation and the objectives achieved with the integration of AI-SPRINT technologies. We analyze how the proposed toolchain effectively addresses a range of challenges and refines processes, discussing its relevance and impact in multiple domains. After a comprehensive overview of the main AI-SPRINT tools used in these scenarios, the paper summarizes of the findings and key lessons learned.

Nico Catalano, Alessandro Maranelli, Agnese Chiatti et al.

Semantic segmentation is a key prerequisite to robust image understanding for applications in \acrlong{ai} and Robotics. \acrlong{fss}, in particular, concerns the extension and optimization of traditional segmentation methods in challenging conditions where limited training examples are available. A predominant approach in \acrlong{fss} is to rely on a single backbone for visual feature extraction. Choosing which backbone to leverage is a deciding factor contributing to the overall performance. In this work, we interrogate on whether fusing features from different backbones can improve the ability of \acrlong{fss} models to capture richer visual features. To tackle this question, we propose and compare two ensembling techniques-Independent Voting and Feature Fusion. Among the available \acrlong{fss} methods, we implement the proposed ensembling techniques on PANet. The module dedicated to predicting segmentation masks from the backbone embeddings in PANet avoids trainable parameters, creating a controlled `in vitro' setting for isolating the impact of different ensembling strategies. Leveraging the complementary strengths of different backbones, our approach outperforms the original single-backbone PANet across standard benchmarks even in challenging one-shot learning scenarios. Specifically, it achieved a performance improvement of +7.37\% on PASCAL-5\textsuperscript{i} and of +10.68\% on COCO-20\textsuperscript{i} in the top-performing scenario where three backbones are combined. These results, together with the qualitative inspection of the predicted subject masks, suggest that relying on multiple backbones in PANet leads to a more comprehensive feature representation, thus expediting the successful application of \acrlong{fss} methods in challenging, data-scarce environments.