Original title: Použití hlubokého učení a unsupervised learning pro analýzu 3D tvarů
Translated title: Unsupervised Deep Learning for 3D Shape Analysis
Authors: Rachler, Ivan ; Kubík, Tibor (referee) ; Španěl, Michal (advisor)
Document type: Master’s theses
Year: 2026
Language: slo
Publisher: Vysoké učení technické v Brně. Fakulta informačních technologií
Abstract: [slo] [eng]
Táto práca sa zaoberá hlbokým učením pre analýzu 3D tvarov s využitím stratégie self-supervised learning. Cieľom je experimentálne overiť prínos tejto stratégie na úlohách klasifikácie a segmentácie častí 3D objektov. Táto práca sa zamerala na architektúru PointMAE, ktorá je založená na maskovanom modelovaní a rekonštrukcii. V rámci experimentov sú aj pokusy o kombináciu tejto architektúry s inou stratégiou self-supervised learning, kontrastívnym učením. V práci sa podarilo ukázať, že model Point-MAE dosiahol až 8,7 % relatívne zlepšenie oproti modelu, ktorý nebol predtrénovaný, na klasifikácií za využitia 5 % dostupných anotovaných trénovacích dát. Na úlohe segmentácie častí sa ukázalo relatívne zlepšenie 9,8 %. Kombinácia PointMAE s kontrastívnym učením neprekonala výsledky základnej architektúry, ale priblížila sa k ním na menej než jedno-percentný relatívny rozdiel. This work deals with deep learning for 3D shape analysis using the self-supervised learning strategy. The aim is to experimentally verify the benefit of this strategy on the tasks of classification and part segmentation of 3D objects. This work focused on the PointMAE architecture, which is based on masked modeling and reconstruction. The experiments also include attempts to combine this architecture with another self-supervised learning strategy, contrastive learning. The work managed to show that the PointMAE model achieved up to an 8.7% relative improvement over a model that was not pretrained, on classification using 5% of the available labeled training data. On the part segmentation task, a relative improvement of 9.8% was shown. The combination of PointMAE with contrastive learning did not surpass the results of the base architecture, but came within less than a one-percent relative difference between them.
Keywords: 3D objects; 3D shape analysis; artificial intelligence; contrastive learning; deep learning; masked modelling; self-supervised learning; unsupervised learning

Institution: Brno University of Technology (web)
Document availability information: Fulltext is available in the Brno University of Technology Digital Library.
Original record: http://hdl.handle.net/11012/260666

Permalink: http://www.nusl.cz/ntk/nusl-781433


The record appears in these collections:
Universities and colleges > Public universities > Brno University of Technology
Academic theses (ETDs) > Master’s theses