Democritus University of Thrace (DUTH)
3-Dimensional triangular meshes: texture processing and retrieva
Abstract
dc:descriptionThe ongoing development of three-dimensional (3D) graphics, coupled with the progress achieved in 3D digitization systems and in combination with the high impact of 3D models' usage in various scientific domains resulted in a significant increase in the available 3D models. One of the most widespread methods of 3D digitization is Structure-from-Motion that estimates 3D structures from two-dimensional image sequences. The efficient organization and management of this type of data in large digital repositories constitute a research challenge. In detail, managing the huge amount of information available in an effective way is imperative and vital and thus classification and content-based retrieval of 3D models constitute an active research field. 3D model classification and content-based retrieval are scientific subfields of multimedia information retrieval offer algorithms for indexing, searching, classification and retrieval of 3D models without the need for textual annotation accompanying each 3D model, that would inevitably lead to doubts and ambiguities due to the diversity of the natural languages.A key motivation of this PhD dissertation was the development of a 3D digitization approach for non-friendly objects for Structure-from-Motion, that is based on the extraction of local features from images using SIFT or SURF algorithms. Thus, if we apply Structure-from-Motion to objects with a surface without features, it is not possible to reconstruct a 3D model. Our method is based on the artificial enrichment of the surface of the object being digitized with feature points through the projection of noise patterns using a projection system. 3D models are often accompanied by texture information, such as artifacts from the Cultural Heritage domain. Their texture can therefore provide useful information about these models (e.g. origin, creator, etc.). In the context of this PhD, we propose a method of transforming 3D model texture maps into a two-dimensional and spatially-consistent form by applying appropriate slices for their transformation into topologically equivalent to the unit-disk and then parametrizing them onto the 2D space. Hence, the implementation of operations from the digital image processing domain, such as edge detection and object recognition is made possible. Moreover, we propose a method of region-of-interest extraction based on the density of the detected SIFT feature points. Furthermore, in a case study, an object was digitized using a multispectral camera and thus using the spatial-consistent texture map representation we are enabled to identify the construction materials of the object utilizing remote sensing and machine learning techniques. We propose two methods of 3D model content-based retrieval and classification using machine learning. In the first case, we rely on both the geometry and the topology of the 3D models, in order to create a feature, the curvature map, which is fed into a convolutional neural network architecture. In particular, the 3D model's surface extremas are identified and in combination with the creation of an appropriate slice the 3D model is parametrized onto the unit-square, where the curvature of the surface embedded in R^3 is rendered. The curvature map is independent of rotation, scaling, and translation. In the second case, we propose a method based on 2D image renderings of the 3D scene, which are fed into a convolutional neural network architecture. In this case, the renderings depend on the position and rotation of the model in the three-dimensional space.Finally, for the evaluation of content-based retrieval algorithms, various metrics are used, either as scalars or graphs. Each one of them highlights a different aspect of the performance of a retrieval method. A thorough definition of the metrics used to evaluate the performance of 3D model content-based retrieval is provided. We developed RETRIEVAL, an online content-based retrieval evaluation tool based on open source technologies. Its functionalities, data input structures, metrics parametrization as well as its pre-installed benchmark dataset are being presented. The goal of RETRIEVAL is to create an environment where each researcher can study and compare the performance of his retrieval algorithms.
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2018
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Ioannakis, Georgios-Alexis
- Ιωαννάκης, Γεώργιος-Αλέξης
Subjects
dc:subject × 27- Υπολογιστική γεωμετρία
- Φωτογραμμετρία
- Μηχανική μάθηση
- Επεξεργασία 3Δ σήματος
- Αναγνώριση προτύπων
- Επεξεργασία 3Δ πλεγματοσειρών
- Ανάκτηση 3Δ αντικειμένων
- 3Δ ανακατασκευή
- Ανάπτυξη λογισμικού
- Computational topology
- Computational geometry
- Photogrammetry
- Machine learning
- 3D signal processing
- Pattern recognition
- 3D mesh processing
- 3D object retrieval
- 3d reconstruction
- Software development
- Φυσικές Επιστήμες
- Μαθηματικά
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Natural Sciences
- Mathematics
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
dc:rights- Statement dc:rights
-
- BY_NC
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/44959