Research Areas
From vision & language to robotics, our research spans the full spectrum of AI. Explore the areas where we push the boundaries of the possible.
Our History
Foundation
Founded by Prof. Rita Cucchiara, initially focused on image processing and computer vision, later expanding to applied AI.
Scientific Growth
Expanded into video understanding, multimedia analytics, and human–machine interaction, becoming one of Italy's leading AI groups with international collaborations.
International Recognition
Selected by Facebook AI Research among 15 EU labs for a strategic GPU partnership.
Joined AIRI
Joined AIRI (formerly Softech), interdepartmental AI center accredited by the Emilia Romagna High Technology Network; initiated the founding collection for the new building; launched AI for health/bioinformatics through the EU projects DeepHealth and Decider.
NVIDIA AI Technology Center
Became host of the NVIDIA AI Technology Center (NVAITC) at the Modena Technopole.
ELLIS Unit Appointment
Appointed as an ELLIS Unit within the European AI network; participated in EU projects (ELISE, ELSA, ELIAS, ELLIOT); promoted the ELLIS Summer School in AI and HPC.
PNRR Projects
Participated in PNRR projects (ITSERR, Ecosister, Fit4MedRob, FAIR as a CNR unit).
EuroHPC
Won the EuroHPC Extreme Scale call for frontier Large-Scale AI research and projects.
Large-Scale Multimodal AI
Our Multimodal AI group develops large-scale architectures that bridge visual perception and natural language understanding, with a focus on trustworthy and deployable AI solutions. We design joint representation models over images, video, and text to support tasks such as visual question answering, dense captioning, cross-modal retrieval, and zero-shot recognition. The group is active within the ELLIS network through EU projects ELIAS, ELSA, and ELLIOT, and collaborates with CINECA on large-scale training infrastructure – including access granted through a competitive win of the EuroHPC Extreme Scale Access Call.
Key research threads include reasoning-augmented and retrieval-augmented generation for knowledge-intensive visual tasks, hallucination mitigation in multimodal large language models via preference optimization, missing-modality robustness in multimodal pipelines, and instruction-guided image edit detection and evaluation. We also advance generative AI for structured visual content, including scalable vector graphics generation through large language models, and inference-time scaling strategies for video diffusion models. The core mission is to push multimodal foundation models toward grounded, interpretable, and reliable reasoning, bridging the gap between research prototypes and real-world deployment across vision-language applications.
Machine Learning & Deep Learning for resilient, online, distributed and incremental models
Our ML/DL group develops the foundational algorithms and training methodologies that power every project in the lab. We design novel neural architectures, study generalisation theory, and build robust training pipelines for both computer-vision, language and time series industrial tasks
Key research threads include continual learning and catastrophic-forgetting mitigation, incremental learning and model updates, neural architecture search (NAS), parameter-efficient fine-tuning (PEFT), model merging and ensemble strategies, self-supervised pretraining on unlabelled data and novel online learning strategies. We also contribute to the reliability and interpretability of deep models, studying calibration, uncertainty estimation, and feature attribution methods. The core mission is to provide technical and theoretical foundation for the resilience, the management and the life cycle of modern deep learning solutions incorporating learning strategies into the model inference and creating modular solutions that can be modified without altering the base model architecture. The studied techniques are the cornerstone of post deployment model resilience allowing to bridge the gap between prototypes and laboratory research and actual real life products.
Medical Imaging
We develop robust and reproducible AI methods for medical imaging, spanning 2D (e.g., ultrasounds and dermoscopy) and 3D volumetric analysis (e.g., CBCT and MRI) and computational pathology. The research combines model design, rigorous benchmarking, and efficient algorithms to enable reliable evaluation and transferability across clinical sites and acquisition settings.
Ongoing projects include automated segmentation in complex 3D scans—such as maxillofacial structures and neurovascular anatomy in CBCT—together with multimodal brain tumour segmentation under missing-modality conditions and large-scale evaluation on community benchmarks (e.g., BraTS). The group also develops learning-based pipelines for ultrasound and other routine clinical modalities, targeting tasks such as detection, grading, and quantitative measurement under real-world acquisition variability. Additional work investigates computer vision methods for clinical analysis of facial behaviour, including automatic detection of facial muscle weakness and neurological conditions through video-based analysis. Clinical applicability is the driving force throughout: methods are designed to be robust across centres and devices, label-efficient, and interpretable, with validation protocols aligned with clinical workflows and regulatory constraints. To foster reproducible research and push research boundaries, the group actively contributes to data collection and public release—organizing international challenges whenever possible—and studies synthetic medical image generation to augment scarce data, probe failure modes, and improve generalization.
Computer Vision
Computer Vision is the heart of AImageLab. Our researchers tackle the full visual understanding stack: low-level image restoration and enhancement, mid-level structural recognition, and high-level semantic scene understanding for images and video in the multimodal AI era.
We are particularly active in action recognition, video object segmentation, pedestrian re-identification, 3D scene reconstruction, and generative modelling and editing (flow matching models, diffusion models, GANs). Our video understanding work targets both efficiency — enabling real-time inference on edge hardware — and accuracy, pushing state-of-the-art on international benchmarks such as ActivityNet, MOT, and Market-1501. We also study flexible multimodal open vocabulary solutions for fundamental computer vision problems stemming from object detection, segmentation, tracking and 3D reconstruction by leveraging the most advanced multimodal open vocabulary architectures and fostering efficient large scale training and modular AI solutions tailored for fine grained vision tasks. Another research direction focuses on modelling human visual attention through eye-movement analysis, including prediction and generative modelling of gaze scanpaths from images and videos, as well as the analysis of egocentric visual data collected with wearable eye-tracking devices.
IoT & Embedded AI
We design end-to-end systems that fuse data from heterogeneous sensor networks — RGB cameras, IMUs, event and depth cameras, LiDARs, and any type of connected sensors — to enable pervasive, context-aware intelligence at the edge.
Our embedded AI research targets microcontroller and AI accelerator deployment of Machine Learning algorithms, tackling the trade-off between model accuracy and resource footprint with quantisation, pruning, and knowledge distillation. Use cases span from Monocular Depth Estimation, precision agriculture, Industrial IoT, and wearable health-sensing platforms.
Robotics & Human-Robot Interaction
We create autonomous systems capable of perceiving, understanding, and acting in complex human environments. Our robots learn from human demonstrations, adapt to novel scenarios through sim-to-real transfer, and communicate intuitively with people. We also develop Embodied AI algorithms that enable agents to navigate and reason in 3D environments, including language-guided navigation where agents follow natural language instructions to reach goals.
Research themes include vision-based manipulation, social navigation with pedestrian intent prediction, multimodal dialogue systems for assistive robotics, and egocentric action anticipation from wearable cameras. A research direction is also the development of Vision-Language-Action (VLA) models that tightly integrate visual perception, language understanding, and motor control into unified architectures for embodied agents. We participate in international competitions and collaborate with industry partners on warehouse automation and collaborative assembly applications.
Bioinformatics and AI for Precision Medicine
We develop AI methods in bioinformatics and computational biology centered on multimodal integration for precision medicine. Our models jointly learn from molecular data—including genomics, gene expression, mutational and proteomic profiles, and three-dimensional structures—together with whole slide imaging features and electronic health records, medical charts, and visit transcripts, enabling therapy response prediction, patient stratification, and advanced therapeutic strategies including ATMPs.
Our work spans bioinformatics and computational biology, combining sequence analysis, molecular network modeling, structural biology representations, and clinical data science within coherent multimodal systems. DNA, RNA, and protein sequences are analyzed to identify regulatory and mutational patterns, while gene expression dynamics, transcriptional regulation, methylation landscapes, protein–protein interaction networks, and three-dimensional structural information are incorporated into system-level representations through reproducible bioinformatics workflows. Quantitative information extracted from whole slide images is integrated as one component of the broader multimodal modeling strategy, alongside molecular and clinical variables. Our frameworks explicitly address high-dimensional, heterogeneous, and partially observed omics and clinical data, including the management of missing modalities and automated information extraction from clinical texts, medical records, and visit transcripts. Through systematic data integration across molecular, imaging, and clinical sources, we aim to enhance patient stratification, quantify and predict therapeutic response, support precision oncology and advanced therapies, and strengthen clinical decision-making processes through interpretable and robust AI methodologies.
Interested in collaborating?
We welcome partnerships with industry, healthcare institutions, and fellow research groups. Reach out to learn how we can work together.