Back to search

Democritus University of Thrace (DUTH)

Unsupervised 3D mesh segmentation

Abstract

dc:description

In this thesis we deal with the problem of unsupervised 3D mesh segmentation. Undoubtedly, 3D mesh segmentation is a difficult and important problem. Semantic analysis of 3D content is strongly related to the parts each object consists of. Due to this reason there is a clear trend in literature towards part-based algorithms. However, current mesh segmentation algorithms are often restricted in a limited range of types of objects, use prior knowledge of the object's class, or human interaction. We review previous approaches and discuss the various existing evaluation frameworks. A novel mesh segmentation algorithm is proposed, which is based on a spectral framework. We introduce a new graph Laplacian based on an heterogeneous graph, which comprises two types of nodes, namely faces and patches. Furthermore, we deal with the problem of partitioning the 3D mesh from the eigenvectors of the aforementioned Laplacian. The proposed approach is motivated by the observation that individual eigenvectors highlight different segments. The nodal sets and nodal domains are used to extract segments from each eigenvector separately, while multiple eigenvectors are combined in a hierarchical manner to produce the final segmentation. Experimental results indicate that the proposed methodology achieves improved performance compared to current unsupervised approaches, while it generates results with error equivalent to the ground truth of the very popular Princeton segmentation benchmark. In addition, we use the segmentation produced by the proposed approach, and show how a part-based representation can be used to retrieve similar objects under the premise that objects of the same class will consist of the same parts. This study further supports the experimental results of the segmentation algorithm. The parts produced for each object are used to construct a graph, where nodes represent parts and adjacency between parts is represented by edges in the graph. A novel matching scheme is proposed, which uses a graph cut algorithm to find a correspondence between parts of one object to the other. The matching scheme penalizes parts (nodes) as well as pairs of parts (edges) without a counterpart. We perform experiments in diverse datasets, which show superior performance of the part-based approach in the case of articulated objects, while being competitive with respect to the state-of-the-art in other datasets.

Degree

thesis:*
Grantor dc:publisher
Democritus University of Thrace (DUTH)
Year dc:date
2016

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Theologou, Panagiotis
  • Θεολόγου, Παναγιώτης

Subjects

dc:subject × 11

Rights

dc:rights
Statement dc:rights
  • BY
Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/40522

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Theologou, Panagiotis; Θεολόγου, Παναγιώτης. Unsupervised 3D mesh segmentation. Democritus University of Thrace (DUTH), 2016. http://hdl.handle.net/10442/hedi/40522