The course overviews visual object tracking approaches, starting with the simple concepts and up to modern deep learning concepts. We will start with single-target visual tracking methods and explore
the paradigm of tracking as detection and correlation filters. We will then continue with early deep-learning methods and more recent methods based on transformer architectures. We will look at
multi-object tracking and explore the role of segmentation in tracking. A part of the course will be dedicated to empirical evaluation, and we will explore the challenges of designing a suitable
evaluation methodology and observe the role of the VOT initiative. The course will conclude with a practical session where we will learn how to evaluate tracking algorithms on a benchmark datasets.
A minimum of 75% attendance is required.
Teaching methods
Lectures, practicum
Assessment methods
90% Project report, 10% oral exam
Bibliography
LUKEŽIČ, Alan, ČEHOVIN ZAJC, Luka, VOJÍŘ, Tomáš, MATAS, Jiří, KRISTAN, Matej. Performance evaluation methodology for long-term single-object tracking. IEEE Transactions on Cybernetics. 2021
ČEHOVIN ZAJC, Luka. A modular toolkit for visual tracking performance evaluation. SoftwareX. 2020
KRISTAN, Matej, LEONARDIS, Aleš, ČEHOVIN ZAJC, Luka, LUKEŽIČ, Alan, DŽUBUR, Benjamin, et al. The Tenth Visual Object Tracking VOT2022 challenge results. ECCV Workshops 2022
KRISTAN, Matej, LEONARDIS, Aleš, ČEHOVIN ZAJC, Luka, LUKEŽIČ, Alan, et al. The Ninth Visual Object Tracking VOT2021 challenge results. ICCVW 2021 : 2021 IEEE/CVF ICCV Workshops
LUKEŽIČ, Alan, ČEHOVIN ZAJC, Luka, VOJÍŘ, Tomáš, MATAS, Jiří, KRISTAN, Matej. FuCoLoT - A Fully-Correlational Long-Term Tracker. ACCV 2018
