Institutional Repository of Vilnius University
Objektų segmentavimas neuroniniais tinklais sprendžiant okliuziją papildytoje realybėje /
Abstract
dc:descriptionThe aim of this paper is to improve the segmentation of human hands in images using the method of solving occlusion in augmented reality that was identified as the most effective. Three methods of solving occlusion -- the problem in which a real object covers a virtual one -- were identified: depth-based, model-based and the usage of neural networks. The model-based method is not suited to dynamic environments and the depth-based method has range restrictions and noise problems. The method using neural networks has the potential to solve arising problems. The occlusion-based application analysis revealed that occlusion solving is not perfect. It is difficult to compare the methods of occlusion solving in practice because software development kits are often utilised and the methods can be combined. Object segmentation is an important part in solving occlusion since a part of the virtual object has to get hidden in a realistic way. Convolutional neural networks are often used for object segmentation. The acquisition of neural network training data sets is a great challenge. The SparseInst neural network architecture and the "EgoHands" data set were chosen and two experiments of training a neural network to segment images of human hands were conducted.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Juodytė, Gabrielė,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD146241377&prefLang=en_US