Πλοήγηση ανά Συγγραφέας "Tzermia, Chrysoula"

Τώρα δείχνει 1 - 2 of 2
  • Μικρογραφία εικόνας
    Τεκμήριο
    Applications of machine learning and object recognition in virtual worlds.
    (ΕΛ.ΜΕ.ΠΑ., ΣΧΟΛΗ ΜΗΧΑΝΙΚΩΝ (ΣΜΗΧ), Τμήμα Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών, 2022-05-11) Tzermia, Chrysoula; Τζερμιά, Χρυσούλα
    In this work, the application of machine learning in the division of objects (models) in the virtual space was studied in order to extract their characteristics and finally their classification and recognition. For this purpose, bibliography and examples of Deep Learning implementation was studied and its application in geometric model recognition were developed.
  • Μικρογραφία εικόνας
    Τεκμήριο
    Sparse point clouds semantic segmentation
    (ΕΛΜΕΠΑ, Σχολή Μηχανικών (ΣΜΗΧ), ΠΜΣ Μηχανικών Πληροφορικής, 2026-06-08) Tzermia, Chrysoula; Τζερμιά, Χρυσούλα; Malamos, Athanasios; Μαλάμος, Αθανάσιος
    This thesis deepens into point clouds semantic annotation and segmentation based on artificial neural network algorithms. According to the literature, the best up to now algorithm that detects directly points of clouds and performs with relatively high efficiency is PointNet. Therefore, our experimental effort focuses on this algorithm and how to increase efficacy. Through data preprocessing, we eliminate point cloud outliers and normalize the coordination system. Additionally, to prevent memory issues we suggest a method to perform data augmentation and by correlating some classes we achieve better accuracy results, preventing data overfitting by adapting PointNet's architecture and learning rate. The evaluation involves comparing the Intersection over Union (IoU) scores of various artificial neural networks utilized in semantic segmentation tasks on outdoor scene point clouds with our customized PointNet version. Additionally, through statistical analysis we present the linkage between efficiency and the dispersion of class related points in point clouds. Finally, identifying incorrect class associations clarifies PointNet's reasoning behind such misclassifications.