Author, Institution: Audrius Kulikajevas, Kaunas University of Technology
Science area, field of science: Natural Sciences, Informatics, N009
Scientific Supervisor: Prof. dr. Rytis Maskeliūnas (Kaunas University of Technology, Technological Sciences, Informatics Engineering, T007)
Dissertation Defence Board of Informatics Science Field:
Prof. Habil. Dr. Rimantas Barauskas (Kaunas University of Technology, Natural Sciences, Informatics, N009) – chairperson
Prof. Habil. Dr. Gintautas Dzemyda (Vilnius University, Natural Sciences, Informatics, N009)
Prof. Dr. Gintaras Palubeckis (Kaunas University of Technology, Natural Sciences, Informatics, N009)
Assoc. Prof. Dr. Agnė Paulauskaitė-Tarasevičienė (Kaunas University of Technology, Natural Sciences, Informatics, N009)
Prof. Habil. Dr. Justyna Patalas-Maliszewska (University of Zielona Gora, Poland, Natural Sciences, Informatics, N009)
The doctoral dissertation is available at the library of Kaunas University of Technology (K. Donelaičio g. 20, Kaunas).
The dissertation defence was held remotely.
Dissertation presents a set of computer models and their application strategy for the completion of a three-dimensional object from a single imperfect depth sensor perspective. The proposed
models can reconstruct the volume and surface of multiple complex and temporally morphing objects per frame from only a single noisy or otherwise distorted depth sensor frame input.
Where the input can be captured by using either LiDAR or structured light depth sensors. With the application of unsupervised deep adversarial auto-refining neural networks the models have
shown to be robust against various types of noise typically seen in real world depth sensors. This is because the novel adversarial-autorefinement branch first removes majority of the distortions
in the depth map, making the input synthetic-like that can be later completed with missing features. Additionally, while the state-of-the-art robustness against noise and reconstruction
quality has a great scientific importance, the conducted research has shown potential commercial applications such as robotics, autonomous vehicles and virtual reality due to their performance in
terms of time being adequate for real-time applications.