Publications

Multimodal Large Language Models (MLLMs) have achieved impressive performance across a variety of vision–language tasks. However, their internal working mechanisms remain largely underexplored. In this work, we introduce FG-TRACER, a framework designed to analyze the information flow between visual and textual modalities in MLLMs in free-form generation. Notably, our numerically stabilized computational method enables the first systematic analysis of multimodal information flow in underexplored domains such as image captioning and chain-of-thought (CoT) reasoning. We apply FG-TRACER to three state-of-the-art MLLMs—LLaVA 1.5, LLaMA 3.2-Vision, and Qwen 2.5-VL—across three vision–language benchmarks—TextVQA, COCO 2014, and ChartQA—and we conduct a word-level analysis of multimodal integration. Our findings uncover distinct patterns of multimodal fusion across models and tasks, demonstrating that fusion dynamics are both model- and task-dependent. Overall, FG-TRACER offers a robust methodology for probing the internal mechanisms of MLLMs in free-form settings, providing new insights into their multimodal reasoning strategies. Our source code is publicly available at https://github.com/AImageLab-zip/FG-TRACER

Large Language Models (LLMs) have revolutionized language modeling through autoregressive generation, enabling strong performance across a wide range of tasks. Recently, Diffusion Language Models (DLMs) have emerged as an alternative paradigm that generates text through iterative denoising rather than next-token prediction, allowing parallel refinement of entire sequences. While numerous diffusion-based architectures have been proposed, differences in evaluation protocols, datasets, inference budgets, and generation hyperparameters make it difficult to compare their capabilities and understand the trade-offs they offer. In this work, we present a systematic experimental analysis of modern DLMs. Specifically, we evaluate eight state-of-the-art DLMs across eight benchmarks spanning reasoning, coding, translation, knowledge, and structured problem solving, while explicitly considering both generation quality and computational efficiency. Beyond downstream evaluation, we analyze the impact of key inference-time factors, including denoising steps, context length, block size, and parallel unmasking strategies, and complement large-scale experiments with controlled comparisons of smaller models trained under identical conditions. Our analysis highlights the strengths and limitations of diffusion-based language modeling across different tasks, architectures, and inference budgets. We show that the behavior of DLMs is strongly influenced by generation-time design choices, leading to distinct trade-offs between performance and computational efficiency. Overall, our study provides practical insights into the capabilities and deployment characteristics of contemporary DLMs.

The PopEYE dataset is a specialized collection of 14,976 near-infrared (NIR) images of the human eye region, specifically designed to support the development and benchmarking of computer vision algorithms for eye-state detection and coarse gaze-direction classification. Each image is provided in a fixed resolution of 772 × 520 pixels in 8-bit grayscale PNG format. The acquisition was performed frontally using a custom-developed Maxwellian-view optical configuration, consisting of a board-level CMOS camera and a specialized lens system where the subject's eye is precisely positioned at the focal point. This setup ensures a high-contrast representation of the anterior segment, making the pupil, iris, limbus, and portions of the sclera and eyelids clearly distinguishable under stable 850 nm infrared illumination. The dataset is categorized into six mutually exclusive classes identified through manual annotation supported by fixed visual aids and an expert system algorithm. The classification includes a correct positioning class for eyes open and properly aligned for clinical measurements (8,160 images), a closed class representing full eye closures such as blinks or sustained lid closure (1,790 images), and four directional classes representing gaze shifts relative to the central optical axis, specifically up (1,379 images), down (1,015 images), left (1,296 images), and right (1,336 images). The data captures the natural anatomical variability of 22 subjects and incorporates common real-world artifacts such as specular reflections from NIR sources and partial pupil occlusions by eyelashes or eyelids. By providing standardized labels and high-resolution NIR imagery, PopEYE serves as a robust resource for training machine learning models intended for real-time patient monitoring during ophthalmic examinations.

In this work, we present TakuNet, a family of ultra-lightweight convolutional neural networks designed for realtime aerial image classification on resource-constrained embedded devices. The proposed TakuNetV2 architecture enhances feature extraction and generalization capabilities through the incorporation of a denser stem coupled with hybrid feature extractor blocks, wherein diverse convolutional operations are synergistically combined to yield richer spatial representations without compromising latency or parameter efficiency. We extensively evaluate the TakuNet family on three public aerial image classification datasets against well-known light-weight and ultra-lightweight architectures, and measured relative performance on five heterogeneous embedded platforms, spanning from CPUs to GPUs, and the Hailo-8 NPU. TakuNet achieves state-of-the-art accuracy and energy efficiency, outperforming competing models in frames per second per additional watt consumed, confirming its suitability for battery-powered edge devices. Additionally, this paper introduces the Astrial platform, highlighting its role in enabling efficient deep learning inference on industrial-grade edge applications. Although current NPU hardware and compiler limitations pose challenges, TakuNet sets a new benchmark for efficient, high-performance embedded artificial intelligence in aerial surveillance and emergency response. Code, models, and weights are publicly available https://github.com/DanielRossi1/TakuNetV2.

Morphing Attack Detection (MAD) systems often suffer from performance degradation when deployed in operational environments, such as airports, that differ from the training domain. We propose an adaptive differential MAD framework that continuously refines a pre-trained detector using live bona fide samples acquired at the gate. The system is memoryless, so no samples are stored in memory to mitigate privacy concerns about the collection of personal data. To prevent the loss of discriminative power caused by bona fide-only adaptation, the method generates synthetic morph samples on-the-fly by combining the current operational subject with identities from an external public or synthetic face dataset. The adaptation process further relies on a quality-aware bona fide selection strategy and a controlled balancing mechanism for synthetic morph generation. Experimental results show that the proposed method improves target-domain specialization while maintaining robustness against morph attacks.

We present a dataset and baseline for ontology-grounded structured prediction from dental Cone-Beam Computed Tomography (CBCT) volumes. Building on the public ToothFairy3 benchmark (532 volumes with expert-level segmentations), we contribute (i) a total of 893 free-text clinical reports for 529 publicly available CBCT volumes, (ii) their conversion into validated RDF/Turtle (Resource Description Framework) instances aligned with a clinician-designed OWL (Web Ontology Language) ontology spanning 13 finding types and multiple qualifier axes, and (iii) a strong baseline demonstrating the effectiveness of our setup and establishing a foundation for future work. We formulate CBCT reporting as a three-stage structured prediction problem—i.e., finding detection, anatomical slot allocation, and property prediction—and introduce a hierarchical evaluation suite of six clinically interpretable metrics that decouple detection, localization, and characterization. A baseline model using frozen multi-scale VoxTell features, a structure-indexed encoder, and ontology-driven prediction heads achieves strong results under 5-fold cross-validation, with stage-decoupled analysis identifying presence detection as the primary deployment bottleneck. Dataset, ontology, and code are publicly released: https://github.com/AImageLab-zip/CBCT-Report

Humanoid robots are increasingly being integrated into diverse scenarios, such as healthcare facilities, social settings, and workplaces. As the need for intuitive control by non-expert users grows, many studies have explored the use of Artificial Intelligence to enable communication and control. However, these approaches are often tailored to specific robots due to the absence of standardized conventions and notation. This study addresses the challenges posed by these inconsistencies and investigates their impact on the ability of Large Language Models (LLMs) to generate accurate 3D robot poses, even when detailed robot specifications are provided as input.

We introduce GazeD, a new 3D gaze estimation method that jointly provides 3D gaze and human pose from a single RGB image. Leveraging the ability of diffusion models to deal with uncertainty, it generates multiple plausible 3D gaze and pose hypotheses based on the 2D context information extracted from the input image. Specifically, we condition the denoising process on the 2D pose, the surroundings of the subject, and the context of the scene. With GazeD we also introduce a novel way of representing the 3D gaze by positioning it as an additional body joint at a fixed distance from the eyes. The rationale is that the gaze is usually closely related to the pose, and thus it can benefit from being jointly denoised during the diffusion process. Evaluations across three benchmark datasets demonstrate that GazeD achieves state-of-the-art performance in 3D gaze estimation, even surpassing methods that rely on temporal information. Project details will be available at https://aimagelab.ing.unimore.it/go/gazed

Atrial fibrillation (AF) is a common cardiac arrhythmia associated with increased risk of stroke and cardiovascular complications. Although electrocardiography (ECG) remains the diagnostic gold standard, remote photoplethysmography (rPPG) has emerged as a promising non-contact approach for extracting pulse-related information from facial videos and enabling unobtrusive rhythm screening. This systematic review investigates the current evidence on facial-video-based rPPG methods for AF detection. Following a PRISMA-based workflow, records were identified from PubMed, Embase, Scopus, IEEE Xplore, and additional search sources. After removing 151 duplicates, 358 records were screened by title and abstract. Of these, 36 full-text reports were assessed for eligibility. Two additional studies, not retrieved by the database search, were identified through reference-list screening, resulting in 19 studies being included in the review. The reviewed literature shows that recent computer-vision and deep learning approaches have improved robustness to motion artifacts, illumination changes, and acquisition noise, often achieving promising detection performance in controlled settings. However, the field remains limited by the small number of available studies, heterogeneous datasets, and scarce clinical validation. Moreover, within the reviewed literature, widespread regulatory-cleared clinical deployment of camera-based rPPG systems for remote AF detection was not identified, indicating that further methodological standardization, larger clinical studies, and regulatory-oriented validation are needed before this technology can become a reliable tool for real-world AF screening.

Visual Autoregressive (VAR) models face a severe "Memory Wall" on edge devices due to large model size and substantial KV-cache requirements. In this work, we analyze the Infinity VAR family (2B and 8B) and propose a compression pipeline for deployment on constrained NVIDIA Jetson systems. We diagnose critical bottlenecks: activation outliers reaching 353x the median and channel-skewed cache variance. To address this, we propose a hybrid pipeline combining SVDQuant—to structurally decouple weight outliers—and Asymmetric Per-Channel KV8 quantization. Our approach reduces the Infinity-8B footprint by 64% (37.1GB →13.3GB), fitting it on the mid-range Orin NX with a 4.1x speedup over Flux.1-dev (W4A4), while achieving superior aesthetic alignment (ImageReward 1.13 vs 0.935). Crucially, we also unlock entry-level feasibility for the Infinity-2B, compressing it from 16.0 to 7.71 GB to enable deployment on the Orin Nano. These results establish a new efficiency standard for high-fidelity generative AI at the edge. The code is available at https://github.com/Henvezz95/deepcompressor.