Keynote Speakers

Visual data analysis is one of the main aims of computer vision and powerful methods have been introduced. This talk will focus on the context of video sports analysis, a framework that shares the same theoretical and algorithmic challenges and that is also receiving a large amount of attention as the beneficiaries of the results include media industries offering broadcast sports games and intelligent live entertainment, professional teams and their coaches in order to analyze both the performance of a particular player and the one of the whole team, and third party companies including those engaged in player scouting and sports event betting. We will describe some problems and proposals aiming to analyze and understand some tactical aspects of a game, including individual and team actions. In particular, methods for detecting and tracking players in soccer and basketball games, for estimating and exploiting player body-orientation for some soccer actions, and for soccer action recognition, will be presented.

The message-passing paradigm has been the “battle horse” of deep learning on graphs for several years, making graph neural networks a big success in a wide range of applications, from particle physics to protein design. From a theoretical viewpoint, it established the link to the Weisfeiler-Lehman hierarchy, allowing to analyse the expressive power of GNNs. We argue that the very “node-and-edge”-centric mindset of current graph deep learning schemes may hinder future progress in the field. As an alternative, we propose physics-inspired “continuous” learning models that open up a new trove of tools from the fields of differential geometry, algebraic topology, and differential equations so far largely unexplored in graph ML.

Physics-inspired learning on graphs

.pdf [21147Kb]

Research in the field of human motion analysis touches on aspects that have an intrinsic multi-level nature. At a low level, there is a need to develop algorithms for estimating the flow field and for detecting features with specific dynamic and semantic characteristics. At a medium level, it is necessary to conceive models to integrate information on wider time intervals and spatial regions: here problems of semantic segmentation, feature tracking and estimation of characteristics on motion primitives are addressed. As we rise to a high level, we face tasks of action/activity recognition and anticipation. All these levels are still addressed individually, although multi-level architectures are emerging. In this speech I will introduce some of these relevant aspects for movement analysis, and then I will present some research directions we are following, citing examples of applications in the medical field and in man-machine interaction.

On the multi-level nature of human motion analysis

.pdf [20341Kb]

Fixed point strategies provide a simplifying and unifying framework to model, analyze, and solve a great variety of problems in image processing. They constitute a natural environment to explain the behavior of advanced convex optimization methods. In addition, an increasing number of problems go beyond optimization since their solutions are not optimal in the classical sense of minimizing a function but, rather, satisfy more general notions of equilibria. Among the formulations that fall outside of the realm of standard minimization methods, we can mention variational inequality and monotone inclusion models, game-theoretic approaches, neural network structures, and plug-and-play methods. This talk will provide an overview of the main tools of fixed point theory and discuss some of their applications to image processing problems. The prominent role played by the forward-backward algorithm will be emphasized.

Fixed point methods in image processing

.pdf [1927Kb]