Σημασιολογικη κατατμηση νεφων σημειων με την βοηθεια της μηχανικης μαθησης

Read full text
Authors: Ilias Tsarpalis
Cited by: 0
References: 15

Abstract

In recent years, neural network-based approaches have emerged as a powerful tool for 3D point cloud semantic segmentation, facilitating a range of engineering and surveying tasks.Despite the potential advantages-automatic labeling of complex scenes and reduced manual effort-practical questions persist about both their performance on different types of data and their user-friendliness for those not deeply versed in programming.This thesis addresses these two core issues through the evaluation of four prominent neural networks-PointNet[1], PointNet++[2], RandLA-Net[3], and PointTransformerV1[4]-on two contrasting datasets, S3DIS[5] and ArCH[6].By applying and evaluating the same models on both datasets, the study highlights how each architecture responds to variations in geometric complexity and data distribution.Key performance indicators, such as mean Intersection-over-Union (mIoU) and per-class accuracy, shed light on how well these networks generalize and what adaptations might be necessary for challenging domains.Beyond raw performance metrics, the work also considers the practicality of these methods for non-programmer engineers.Many in the fields of surveying, construction, or heritage preservation may not have extensive coding experience, yet they stand to benefit significantly from automatic point cloud segmentation.Through hands-on trials and process documentation, the thesis assesses whether these networks can be installed, configured, and deployed with minimal specialized knowledge.The findings aim to guide both researchers and practitioners in choosing not only the most accurate neural network for a given 3D segmentation task, but also the most user-friendly approach-offering insights into the potential for wider adoption in everyday engineering workflows. V ΕυχαριστίεςΟλοκληρώνοντας τη μεταπτυχιακή μου εργασία, θα ήθελα να ευχαριστήσω όσους με βοήθησαν και με στήριξαν όλο αυτό το διάστημα.Αρχικά θα ήθελα να ευχαριστήσω τον επιβλέποντά μου και ομότιμο καθηγητή Ε.Μ.Π.κ.Ανδρέα Γεωργόπουλο για την ανάθεση του συγκεκριμένου θέματος, τη συνεχή καθοδήγησή του, το ενδιαφέρον του αλλά και τη βοήθειά του σε θέματα γνωστικού αντικειμένου Ιδιαίτερες ευχαριστίες θα ήθελα να απευθύνω στον υποψήφιο διδάκτορα Ε.Μ.Π.Θοδωρή Μπέτσα, για τις συμβουλές του, την άψογη συνεργασία που έχουμε, για τον πολύτιμο χρόνο που αφιέρωσε και την βοήθεια του στην ολοκλήρωση της μεταπτυχιακής μου εργασίας.

Fields of Study & Subjects

Advanced Decision-Making Techniques Remote-Sensing Image Classification Data Mining and Machine Learning Applications

Publication details

Journal / source
DSpace - NTUA (National Technical University of Athens)
Publication date
1 January 2026
Language
EL