Limited Data in Biometrics: Challenges and Emerging Directions
Guido Borghi
 University of Modena and Reggio Emilia, Italy
Abstract:
The development of effective biometric systems increasingly relies on large, diverse, and well-annotated datasets. However, in many real-world scenarios, data scarcity remains a major challenge. Collecting biometric data is often expensive, time-consuming, and subject to strict privacy regulations that limit its storage, sharing, and reuse. Moreover, available datasets frequently lack diversity in terms of ethnicity, age, and other demographic variables, raising concerns about bias and fairness. These challenges are particularly critical in deep learning approaches, which typically require large amounts of high-quality annotated data - often labelled by domain experts, whose availability is limited. This workshop will explore the key challenges in developing biometric systems, specifically Morphing Attack Detection and Face Image Quality Assessment systems, with limited data and discuss promising directions to address them. In particular, we will examine the potential of generative models to synthesize realistic biometric data, and investigate alternative learning paradigms such as Continual Learning.
Bio:
Guido Borghi is an Associate Professor at the University of Modena and Reggio Emilia, Italy. Previously, he was an Assistant Professor at the University of Bologna. His research topics mainly belong to Computer Vision and Artificial Intelligence. In particular, he works in the field of Face Analysis (Face Recognition, Head Pose Estimation, Quality Assessment), Natural User Interfaces (Gesture and Activity Recognition), applied to Biometrics (Morphing Attack Detection), Robotics (Robot Pose Estimation) and Automotive (Driver Monitoring). He is Associate Editor of Simulation & Gaming and Guest Editor of CVIU. He co-organized several workshops (T-CAP@ICIAP, ICPR, ECCV, and WCPA@ECCV) and a tutorial at IJCB in the field of human and face analysis. He regularly serves as a reviewer for international conferences and journals, such as TPAMI, TIST, TBIOM, CVPR, ECCV and ICCV.
Multi-Task Learning for Gesture Recognition: Addressing Human Movement Complexity under Data Scarcity
Gavriela Senteri
Mines Paris-PSL, France
Abstract:
Understanding human movement, especially in professional, social, or embodied settings, requires more than recognizing patterns in pixels or postures. Movements are not only performed but are also situated: shaped by task, context, dexterity, skills. Yet, in many real-world applications, annotated data is scarce, task boundaries are fuzzy, and gesture executions vary across individuals, time, and environment. This makes gesture recognition a particularly rich but challenging field for learning systems. While many machine learning models achieve great results on curated benchmarks, their assumptions often fail in real-life gesture recognition scenarios. Unlike static images or bounded classification tasks, gestures unfold in time, and vary not only among different people but even within the same one. This gap sets the need for approaches that are both data-efficient and structurally adaptable. This talk explores how structured representations of movement and adaptive learning strategies can support movement recognition under such constraints. Drawing on our recent work in motion analysis, and hierarchical modeling, the discussion will center on the need to align Machine Learning methods with the layered, temporal nature of human gestures that are able to learn from what is shared and adapting to what is unique.
Bio:
Gavriela Senteri is a researcher with a background in Artificial Intelligence for human-computer interaction, movement recognition, and analysis. She is currently pursuing a PhD at Mines Paris - PSL on the "Exploration of Meta-Learning and Multi-Task Learning for the hierarchical and simultaneous recognition of actions, activities, and intentions in professional environments". She has worked on many EU (Horizon, FP7, Creative Europe) projects, focusing on gesture recognition algorithms for human-robot collaboration, sensorimotor feedback systems, and applications that enhance the Cultural and Creative Industries through their digitization. She has also received a B.S. degree in Applied Informatics from the University of Macedonia, and an M.S. degree in Computational intelligence and Digital Media from the Aristotle University of Thessaloniki. Her general research is focused on the analysis of vision and sound modules for human movement analysis and the recognition of expert professionals in the handcraft and manufacturing industries.