Publish Date: 23.12.2022
Category: Outstanding research achievements, Interdisciplinary research, Our contribution to sustainable development goals
Sustainable development goals: 9 Industry, innovation and infrastructure, 11 Sustainable cities and communities (Indicators)
Visual tracking of objects is one of the central issues of computer vision with major application potential in ensuring the safety of autonomous robots and vehicles, video editors, the entertainment industry, 3D reconstruction and enhanced reality. It deals with a situation where the object is selected only in the first image, and the tracking algorithm must estimate its position in all subsequent images, and adapt to changes in its appearance and concealment by other objects. The traditional methods are capable of providing the location only as a rectangular region, which is complex but not sufficiently precise for many applications.
We proposed the new D3S tracker, which is capable of calculating an extremely precise model of the object and segmenting it from the background in each image. This represents a new shift in trackers based on deep learning. We published the D3S in IEEE TPAMI, which with an IF > 24 is the most prestigious artificial intelligence (AI) journal, and presented a preliminary version at the CVPR conference (together they received more than 150 citations). The good reception of the D3S is also evidenced by the fact that plenty of the best trackers for the international VOT2021 challenge were based precisely on the concepts represented in the D3S. In addition to the tracker we also published a new methodology and currently the most complex database, which has become part of the standard evaluation of long-term trackers in the field of computer vision (VOT). The methodology is published in the journal IEEE TCyb, which with an influence factor of 19.118 is the foremost in two technical fields.
Figure 1: Architecture of D3S (left) and examples of localisation of objects (right).
Authors:
Asist. dr. Alan Lukežič, doc. dr. Luka Čehovin Zajc, prof. dr. Matej Kristan.
Sources:
Lukežič, A., Matas, J., Kristan, M. (2022). A discriminative single-shot segmentation network for visual object tracking. IEEE transactions on pattern analysis and machine intelligence, 44(12), 9742–9755. https://doi.org/10.1109/
TPAMI.2021.3137933, IF(2021)=24,314, (2/145 Computer science, Artificial intelligence), A'', h-5 indeks: 165 [COBISS.SI-ID 94955523]
Lukežič, A., Čehovin Zajc, L., Vojíř, T., Matas, J., Kristan, M. (2021). Performance evaluation methodology for long-term single-object tracking. IEEE transactions on cybernetics, 51(12), 6305–6318. https://doi.org/10.1109/TCYB.2020.2980618, IF(2021)=19,118, (3/145 Computer science, Artificial intelligence), A'', h-5 indeks: 142 [COBISS.SI-ID 1538564803]