PhD Researcher at University of Udine
Matteo Dunnhofer is a postdoctoral researcher at the Machine Learning and Perception Lab of the University of Udine, Italy. From the same institute, he received the BSc and MSc in Computer Science in 2016 and 2018, and the PhD in Industrial and Information Engineering in 2022. In 2018, he has been visiting student at the Australian Centre for Robotic Vision hosted at the Queensland University of Technology, Australia.
His research focuses on the use of deep learning to tackle computer vision problems, especially for visual object tracking. He is also interested in applying deep learning techniques in the context of medical image analysis and video-based sport analytics. On these topics, he published several papers that appeared in international journals and conferences, and he organized workshops and tutorials. In 2021, he has been awarded for winning the Visual Object Tracking VOT2021 Long-term Challenge.
Tuesday, July 5th – Academic participants
Deep Learning for Visual Object Tracking
In its simplest definition, visual object tracking consists in the persistent recognition and localization of a generic target object in a video. Several challenges such as object occlusions, pose and scale changes, rotations and shape variations, and the presence of similar objects, must be tackled to accurately keep track of a target’s position. The ultimate goal of generic object tracking is to build robust models capable to overcome such challenging factors. In the past, such issues have been addressed by disparate principles formalizing the concepts of appearance model, motion model, and matching operation. In recent years, algorithms based on deep learning tried to learn such conceptual blocks by exploiting the ability of deep neural networks in learning complex functions from visual examples. Thanks to these advancements, today deep learning-based solutions are the way-to-go to implement strong visual tracking algorithms. The goal of this laboratory is to present the latest progress in the exploitation of deep learning for building an accurate visual tracker. After the introduction of the fundamental concepts of the visual object tracking domain, the session will describe how the state-of-the-art solutions employ deep learning architectures and optimization techniques. The tutorial will also cover the datasets, protocols and metrics available to evaluate deep learning-based trackers, as well as the most popular software tools developed by the community.
Wednesday, July 6th – Business participants
Deep Reinforcement Learning for Robot Control
Reinforcement Learning (RL) is a promising paradigm to optimise machine learning models with limited supervision. In recent years, a successful revival of such techniques took place in the research community thanks to the demonstration of their effectiveness for the optimisation of deep neural networks. Among the many applications currently benefiting RL methods, there is robot control. The basic idea is model the controller of a robot with a neural network, define an high-level goal that the robot should achieve, and use RL to autonomously learn a control behaviour that makes the robot fulfil the objective. This lab lecture will deal this problem. A neural network will be trained via state-of-the-art RL optimisation algorithms to provide the control parameters of a robotic arm that must pursue some simple goals in a simulated environment. After the presentation of the fundamental concepts to reason about RL problems, the tutorial will show how to use the most recent open-source software tools to solve the problem of interest.