In precision viticulture, the early identification of phytosanitary threats and accurate yield estimation are critical for optimizing pesticide use and ensuring economic sustainability, wine safety and wine quality. This study presents a comprehensive IoT based framework for the autonomous monitoring of vineyards, based on a Raspberry PI single board computer for on-tractor deployment. We evaluate State-Of-The-Art (SOTA) object detection architectures, including YOLOv11m, YOLOv26m, and RF-DETR nano and small-size, to identify a broad spectrum and impactful of grapevine pathologies: Phylloxera, Flavescence Dorée, Grey mould, Black Rot, Downy Mildew, Powdery Mildew and Esca disease. The best model was the RF-DETR small size model achieved overall values of precision and recall of 0.75 and 0.58, outperforming the YOLOv11m, Yolov26m and RF-DETR nano size model. The results of the inference carried out on real-world images are used, together with geospatial information, to develop a georeferenced map of the yield and diseases ready for GIS and agronomic management use.

The deployment of Vision-Language Models (VLMs) in critical domains like disaster management requires high-quality multimodal datasets, especially for transferring knowledge via Data-Free Knowledge Distillation (DFKD). However, existing datasets in this domain either entirely lack descriptive text, such as Incidents1M, or suffer from severe text-image semantic misalignment, such as CrisisMMD. In this work, we present a novel methodology to construct and automatically validate a large-scale multimodal dataset for disaster response. Starting from the vision-only Incidents1M, we successfully recovered 100,000 images and generated high-fidelity textual descriptions using two distinct Qwen3.5 architectures: a 4B dense model and a 35B Mixture-of-Experts (MoE) model. To ensure the generated captions provide reliable semantic anchoring for DFKD, we introduce an image-blind LLM-as-a-Judge validation pipeline leveraging Qwen3.5-9B. By intentionally obscuring the original image from the judge, this evaluator accurately simulates the modality gap of the student model during data-free distillation. Our evaluation across 173,179 label pairs demonstrates a high semantic agreement (78.65/100) between the two architectures. Furthermore, the automated evaluation reveals a conservative captioning behaviour, characterized by a high Precision (77.6

Heritage 3D Reconstruction is a digital bridge to the past, providing a vivid window into history by meticulously preserving the intricate details of cultural artifacts and sites and bringing them to life for future generations to explore and appreciate. In recent years, novel Neural Rendering methods, such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), have shown promising results for 3D reconstruction. Despite their rapid development, a systematic comparison of these approaches in the context of 360^∘ imagery, particularly within Digital Cultural Heritage (DCH), remains largely unexplored. In this work, we present a novel comparative framework specifically designed to address this gap, performing a comprehensive assessment of traditional 360^∘ photogrammetry against emerging Neural Rendering approaches. To validate its effectiveness, we conducted a comparative analysis on two different Cultural Heritage scenarios: a large-scale outdoor environment and an object-centered setting. These case studies, which present a variety of challenges in the DCH domain, were selected to evaluate the ability of photogrammetry, NeRF, and 3DGS-based methods to reconstruct and render complex scenes in terms of both geometric accuracy and visual quality. The results highlight the differences between Neural Rendering approaches and conventional methods, particularly in handling spherical geometry, adaptability to varying lighting conditions and efficiency in data acquisition and processing. These findings confirm the continued robustness of spherical photogrammetry for metric reconstruction, while also underscoring the potential of Neural Rendering-based scene representation methods as promising tools for DCH preservation.

Convolutional Kolmogorov–Arnold Networks (KANs) replace the fixed weights of a convolutional kernel with learnable univariate functions. The dominant formulation attaches one such function to every kernel entry and lets it act on pixel values, expressive but parameter-heavy and prone to overfitting. We argue that the learnable functions are better placed in the structure of the convolution than on each edge, and we organise the design space along a single axis: whether the function acts on the pixel values or on the filter shape. We study three realisations. SV-KAN applies one shared univariate function to the values and leaves the spatial filter free and static, aa classical convolution with a single learnable shared activation. AG-KAN keeps the shared value function but supplies the spatial structure through a content-adaptive Gaussian gate. RF-KAN instead moves the learnable functions onto the filter shape, building each filter from oriented ridge profiles expanded in a localised oscillatory (Morlet) wavelet basis with content-adaptive amplitudes. Under a matched four-layer protocol with in-run references and three seeds, RF-KAN and SV-KAN reach 88.47±0.10% and 88.20±0.31% on CIFAR-10 and 64.40±0.19% and 64.57±0.30% on CIFAR-100, at about 0.4M parameters. At this matched scale the shape model and the simplest value model meet at the top, both above a plain convolution and every per-edge KAN we tested, including the official Gram variant, at roughly a fifth of the parameters. A controlled study attributes the RF-KAN gain to an intrinsically localised oscillatory basis and to content adaptivity, and an ablation that removes the learned shape entirely, leaving only the shared value function, collapses accuracy by over forty points, identifying the learned shape as the load-bearing ingredient at this scale.

Aerial surveillance using Unmanned Aerial Vehicles (UAVs) faces challenges in detecting small targets, particularly under varying illumination conditions. Single-modality detectors often degrade in performance when targets are visually camouflaged or when lighting is poor. To improve robustness across day/night and reduced-visibility scenarios, this paper introduces multimodal RGB-thermal perception, leveraging the complementary information provided by visible-spectrum and thermal infrared imagery. An enhanced multimodal architecture based on DEYOLO (Dual-Feature-Enhancement YOLO) is proposed, specifically optimized for small object detection in medium-high altitude imagery. The model employs dual RGB and thermal backbones and integrates lightweight architectural refinements aimed at improving multi-scale representation and cross-modal alignment. In particular, the feature pyramid is extended toward higher-resolution levels to better capture small targets, with SPDConv introduced as an additional module in the backbone. The neck is enhanced with SPANet to improve the extraction of fine-grained details, while attention-based fusion modules adaptively weight spatial and channel information across modalities. The approach is evaluated on both a public RGB-thermal UAV dataset and a custom dual-sensor aerial dataset. Experimental results show that the proposed multimodal configuration achieves a $78 \% \text{mAP}_{50}$, improving detection accuracy and robustness compared to single-modality baselines, especially in low-light and cluttered environments. Ablation studies further confirm that each architectural component contributes complementary improvements, especially in enhancing multi-scale representation and cross-modal feature fusion.

This study provides the first comprehensive evaluation of foundation models in fetal ultrasound (US) imaging under low inter-class variability conditions. While recent vision foundation models such as DINOv3 have shown remarkable transferability across medical domains, their ability to discriminate anatomically similar structures has not been systematically investigated. We address this gap by focusing on fetal brain standard planes–transthalamic (TT), transventricular (TV), and transcerebellar (TC)–which exhibit highly overlapping anatomical features and pose a critical challenge for reliable biometric assessment. To ensure a fair and reproducible evaluation, all publicly available fetal ultrasound datasets were curated and aggregated into a unified multicenter benchmark, FetalUS-188K, comprising more than 188,000 annotated images from heterogeneous acquisition settings. DINOv3 was pretrained in a self-supervised manner to learn ultrasound-aware representations. The learned features were then evaluated through standardized adaptation protocols, including linear probing with frozen backbone and full fine-tuning, under two initialization schemes: (i) pretraining on FetalUS-188K and (ii) initialization from natural-image DINOv3 weights. Models pretrained on fetal ultrasound data consistently outperformed those initialized on natural images, yielding weighted F1-score improvements of up to 21 ≤ 16

This paper presents an automated fish detection pipeline for ROV-acquired underwater imagery collected in the Mediterranean Sea. The task poses significant challenges due to low image quality, high inter-scene variability in water turbidity and depth, and the presence of visually confounding elements such as algae and offshore platform structures. We adopt the YOLOv8 architecture in both its Small and Large variants and evaluate four training configurations under a 5-fold cross-validation scheme. To address the limitations of the default Ultralytics checkpoint selection strategy, we introduce a custom weighted scoring function that jointly optimizes precision, recall, mAP50, and validation losses, promoting more balanced and generalizable model selection. The best-performing configuration—YOLOv8-Small with online data augmentation, trained for 250 epochs—achieves a precision of 0.639, recall of 0.599, F1 score of 0.616, and mAP50 of 0.574, while limiting training time to 41 hours and carbon footprint to 3.90 kg CO2eq. These results demonstrate the viability of the proposed pipeline for large-scale ecological monitoring, while highlighting the need for further dataset refinement to address persistent annotation noise and image quality constraints.

Recent advances in Extended Reality (XR) allow the creation of immersive fitness and virtual coaching systems. However, it remains unclear whether immersive visualization modalities influence users’ perception, understanding, and confidence in performing exercises. Moreover, it is unclear whether this improves those with respect to a classical 2D visualization. In this work, we investigate the role of a parametric 3D human model in enhancing workout perception and social presence within immersive fitness environments. To this date, we have designed a modular XR system, named XaRNold, that allows the visualization of single/multi-view video, and a parametric 3D model called Skinned Multi-Person Linear Model (SMPL) showing how to perform a physical exercise. We conducted a controlled user study (N=30) to compare participants’ cognitive load, understanding, technology acceptance, usability, and social perceptions and engagement across visualization modes. We employ the Fit3D dataset, which provides aligned 2D, multi-view, and related 3D poses data for diverse physical exercises. The results indicate that a parametric virtual human model provides a favorable trade-off between the measured constructs, offering a clearer understanding of why and how avatar-based guidance can be used and paving the way for more effective and empathic virtual fitness systems.

The rapid adoption of Large Language Models (LLMs) in interactive systems has enabled the creation of dynamic, open-ended Role-Playing Agents (RPAs). However, evaluating these agents remains a significant challenge, as standard NLP metrics fail to capture the nuances of role adherence, logical consistency, and long-term narrative stability. This paper introduces RPA-Check, a multi-stage automated evaluation framework designed to objectively assess the performance of LLM-based RPAs in complex, constraints-heavy environments. Our methodology is based on a four-step pipeline: (1) Dimension Definition, establishing high-level qualitative behavioral criteria; (2) Augmentation, where these requirements are expanded into granular boolean checklist indicators; (3) Semantic Filtering, to ensure indicator objectivity, no redundancy and agent isolation; and (4) LLM-as-a-Judge Evaluation, which employs chain-of-thought verification to score agent fidelity. We validate this framework by applying it to LLM Court, a serious game for forensic training involving several quantized local models. Experimental results across five distinct legal scenarios demonstrate the framework's ability to identify subtle trade-offs between model size, reasoning depth, and operational stability. Notably, the findings reveal an inverse relationship between parametric scale and procedural consistency, showing that smaller, adequately instruction-tuned models (8-9B) can outperform larger architectures prone to user-alignment bias or sycophancy. RPA-Check thus provides a standardized and reproducible metric for future research in generative agent evaluation within specialized domains.

Effectively and precisely identifying damages after a seismic event is fundamental to ensuring timely intervention and optimizing the building life cycle. The combination of multisensor inspection and artificial intelligence (AI) technologies is powerful in this context. This work presents the results of the development of a multisensor measurement procedure relying on drone-embedded visible and thermal cameras and a terrestrial laser scanner (TLS) to acquire data on a masonry church damaged by an earthquake. These data were carefully aligned into a unified reference system and enabled the reconstruction of high-resolution 3-D models of the building, from which multichannel orthophotos were extracted, with visible, thermal, and depth data, to identify surface lesions. The analysis procedure involved three main stages: 1) preprocessing using principal component analysis (PCAs) to improve feature separability and reduce redundancy between channels; 2) lesion detection using a deep learning U-Net segmentation model trained to identify surface cracks; and 3) morphological postprocessing to refine the predicted masks and eliminate (or reduce) false positives predictions. The combination of multimodal data improves lesion detection, highlighting cracks that are not immediately visible, particularly when thermal gradients or geometric discontinuities provided complementary evidence. The results are promising and demonstrate both the feasibility of lesion identification and the validity of the proposed approach in the context of structural health monitoring. The use of sensors that can acquire data remotely and without contact offers significant advantages in terms of safety in a postearthquake context.

In the domain of computer graphics, achieving high visual quality in real-time rendering remains a formidable challenge due to the inherent time-quality trade-off. Conventional real-time rendering engines sacrifice visual fidelity for interactive performance, while image generation using path-tracing techniques can be exceedingly time-consuming. In this paper, we introduce RenderGAN, a deep learning-based solution designed to address this critical challenge in real-time rendering. RenderGAN uses G-Buffers and information from a real-time rendering engine as inputs to produce output images with exceptional visual fidelity. Its encoder-decoder architecture, trained using the Generative Adversarial Network (GAN) framework with perceptual loss, enhances image realism. To evaluate RenderGAN's effectiveness, we quantitatively compare the generated images with those of a path-tracing engine, obtaining a remarkable Universal Image Quality Index (UIQI) value of 0.898. RenderGAN's open-source nature fosters collaboration, driving advancements in real-time computer graphics and rendering techniques. By bridging the gap between real-time and path-tracing rendering, RenderGAN opens new horizons for accelerated image generation, inspiring innovation and unlocking the full potential of real-time visual experiences. Project page: https://github.com/marcomameli1992/RenderNet

In an age where consumer preferences are as diverse as they are dynamic, the ability to offer personalized fashion recommendations at scale remains a significant challenge for retailers. Consumers seek a shopping experience that not only understands their unique style preferences but also dynamically adapts to their evolving tastes. The fashion industry is at a crossroads, facing increasing consumer demand for personalization, sustainability and transparency in a rapidly evolving digital marketplace. Traditional retail practices, while rich in tradition and artistry, often struggle to up-to-date with the rapidly, ethically-conscious and technology-driven expectations of today’s consumers. “OutfitAI” is designed to address these challenges by leveraging the power of deep learning to revolutionize the fashion retail experience. By automating the process of background removal in fashion images, using advanced algorithms for personalized product matching, and integrating sustainability filters into the product discovery process, OutfitAI aims to deliver a shopping experience that is not only personalized and engaging, but also aligned with the ethical and environmental values of the contemporary consumer. Unlike existing solutions, OutfitAI uses state-of-the-art semantic segmentation for precise background removal, enabling detailed feature extraction from fashion images. This process enables accurate matching of user-uploaded images with similar fashion items from an extensive database of eco-friendly and ethically produced products sourced from leading e-tailers. Setting itself apart from the current state of the art, OutfitAI places a strong emphasis on ethical data use and privacy, implementing robust measures to ensure user privacy and transparency. It also pioneers the integration of sustainability into the digital fashion discovery process, promoting responsible consumption patterns among users. Through a comprehensive system architecture that combines technical innovation with a commitment to ethics and sustainability, OutfitAI not only addresses the technological needs of the fashion retail industry, but also responds to the growing demand for more responsible and transparent consumer technologies.

The rapid advancement of Artificial Intelligence (AI) is transforming the construction sector, particularly in site monitoring and safety management. Real-time monitoring enables the automatic detection of work progress issues, anomalies, and hazardous situations. However, no existing Deep Learning (DL)-based system is specifically designed to utilize Unmanned Aerial Vehicles (UAVs) for excavation area monitoring. This study presents an automated workflow that integrates UAV imagery with DL architectures, featuring a 1D Convolutional Neural Network (1D-CNN) for classifying excavation work phases and a VGG16 network for detecting safety fences. These technologies are incorporated into a Decision Support System (DSS), which automates report generation and enhances decision-making by providing structured, data-driven insights. The system was validated in a real-world case study involving an oil and gas construction company, demonstrating its ability to streamline site management tasks and improve safety oversight. Compared to traditional monitoring methods, our approach leverages UAV technology and DL methodologies to provide higher accuracy, efficiency, and scalability in excavation site monitoring. This contribution supports the digital transformation of construction management, offering a practical and innovative solution for real-time progress tracking and compliance verification.

Recent advances in Novel-View Synthesis (NVS) and 3D Generation (3DGen) from 2D images have marked significant progress in various domains. While the Structure-from-Motion (SfM) and Multi-View Stereo (MVS) pipelines remain prevalent, their limitations have driven the exploration of Deep Learning (DL)-based methods. Among these, Neural Radiance Fields (NeRFs) stand out for their exceptional capabilities in novel view synthesis and 3D reconstruction. However, their reliance on large, diverse 2D images for training, which capture the same scene from different perspectives, poses challenges. To address these challenges, our research proposes a module that introduces innovative data-centric strategies to improve the fidelity of novel view synthesis and reconstruction of NeRFs. In particular, the adopted strategy relies on depth priors, RGB masks, geometrical warping, and deep learning-based image restoration to improve the training and performance of NeRF models, following a human-in-the-loop approach. This module paves the way for a novel data-centric and DL-driven, to improve performances in NeRFs, which is adaptable across different NeRF architectures. Through a comprehensive quantitative-qualitative analysis of such a framework, on a challenging NeRF benchmark dataset, we demonstrate the effectiveness and versatility of our approach.

The increasing global water scarcity has made the safe reuse of treated wastewater essential, especially in agriculture, where untreated water poses risks to public health. Digitalizing Wastewater Treatment Plants (WWTPs) can enhance real-time water quality monitoring and optimize plant operations. This study implements an Early Warning System (EWS) at the Peschiera Borromeo WWTP in Milan, Italy, using predictive models based on simulated and real datasets to estimate key water quality parameters like Chemical Oxygen Demand (COD) and Total Suspended Solids (TSS). A Multi-Task Learning (MTL) neural network provided real-time predictions and sensor malfunction detection, while a Long Short-Term Memory (LSTM) network forecasted water quality up to six hours ahead. Simulated data showed high correlation coefficients above 0.98, but real-world data reduced performance to 0.31-0.67. Despite this, the EWS shows strong potential for improving treated water reuse reliability and operational efficiency in WWTPs.

The translation of images into detailed three-dimensional (3D) models represents a critical challenge in digital content creation, particularly for the Creative Industries (CI). Traditional 3D modeling methods are resource-intensive, while recent advances in Neural Rendering (NR) have introduced efficient and automated solutions. In the fashion industry, where visual fidelity and rapid prototyping are crucial, NR techniques such as Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (3DGS) are becoming relevant resources to digitize complex geometries and textures with a low-cost approach, enabling different applications like virtual fitting rooms, immersive e-commerce, and digital prototyping. However, the evaluation of their effectiveness in this field is still in its early phases. To address the industry’s needs, we propose Fashion immersive Neural Rendering Interface (FENRI), a novel framework that integrates NR techniques for reconstructing 3D models from 2D images of fashion items. FENRI includes a WebXR-based visualization platform that allows immersive comparison and evaluation of NR-generated 3D models, supporting both experts and non-experts in selecting optimal designs. We applied FENRI to footwear design, collecting a novel dataset to compare NeRF and 3DGS methods through quantitative and qualitative analyses. In our study, the 3DGS method demonstrated superior performance over NeRF, as highlighted by the higher Peak Signal-to-Noise Ratio (PSNR) values (37.65 vs. 29.03, respectively) and Structural Similarity Index Measure (SSIM) scores (0.99 vs 0.96), while exhibiting a lower Learned Perceptual Image Patch Similarity (LPIPS) values (0.01 vs 0.04). Moreover, we showed that, by applying classical mesh post-processing techniques, we can increase the topological and visual quality of the 3D models synthesized by NR methods. These findings highlight the potential of NR techniques in the fashion industry’s digital pipeline. By enabling rapid, immersive, and visually compelling design iterations, FENRI offers a scalable solution to the fashion industry’s demand for high-quality 3D reconstructions, promoting innovation and sustainability.

Can machine learning transform the development of marketing personas for the luxury fashion sector? In the domain of luxury fashion, the potential of machine learning and Artificial Intelligence (AI) has only begun to be explored. While the transformative capabilities of AI are widely acknowledged across various sectors, the luxury fashion marketing domain presents a unique opportunity for groundbreaking advancements. This research demonstrates the application of AI algorithms on a comprehensive dataset that includes transactional, demographic, and behavioural data for a leading global luxury fashion company in Italy. The study concentrates not only on the technical abilities of the algorithms but also on their practical effects, providing beneficial insights for luxury fashion marketers, brand managers, and decision-makers. This study serves as evidence of the effectiveness of machine learning in fashion marketing, revealing its potential for transforming the luxury fashion industry. By extensive datasets, the study demonstrates how machine learning can offer a detailed insight into consumer behaviour, needs, and preferences. With luxury fashion synonymous with exclusivity and uniqueness, these insights can be invaluable in tailoring marketing strategies that resonate with discerning clients. The methodology and findings of the study are set to be a benchmark for future explorations in this domain.

Vegetation restoration is crucial for environmental conservation and maintaining ecosystem services. Traditional methods, such as manual inspections and expert photo interpretation, have been widely used to assess vegetation recovery but are labor-intensive, time-consuming, and prone to human bias. In contrast, modern Artificial Intelligence (AI) based methods use satellite imagery for efficient vegetation analysis, enabling largescale monitoring with minimal human effort. This paper introduces VegRecoverAI, a comprehensive system that leverages multisource satellite data from Landsat, Sentinel-2, and PlanetScope. VegRecoverAI autonomously detects both subtle and significant vegetation changes, providing a reliable alternative to manual assessment. The system extracts NDVI time series data, detects vegetation change and uses an ensemble of forecasting models to predict future vegetation restoration. The system is demonstrated as a case study following gas pipeline construction in Italy. The results indicate that VegRecoverAI is automated and a scalable solution complementary to traditional techniques to support proactive environmental management.