Publicação

Optimization of geometry and texture for 3D reconstruction using RGB-D data

Ver documento

Detalhes bibliográficos
Resumo:Three-dimensional (3D) reconstruction is the creation of 3D models from the captured shape and appearance of real objects. It is a long term topic of investigation with numerous systems and algorithms, having gained importance in various areas, such as architecture, robotics, autonomous driving, medicine, agriculture, and archaeology. The current state of 3D reconstruction of real-world scenes is still not adequate for many applications that require photorealistic quality, including Virtual Reality (VR) and Augmented Reality (AR) experiences, and other human-centered applications. Furthermore, the best current solutions are very expensive, presenting high capital and logistical costs. This PhD aims to study and propose methods for enhancing geometry and texture in 3D reconstructions, resulting in complete virtual models with photorealistic quality. To this end we present a novel method for mirror segmentation and pose estimation, tackling artefacts introduced by reflective surfaces, a neural-based colour correction technique to improve texture consistency across multiple views, mitigating the appearance of texture seams, and introduce objective metrics to quantitatively assess the visual quality of textured 3D meshes. Extensive experiments validate the effectiveness of these contributions, demonstrating significant improvements in geometric accuracy and visual fidelity of the produced models.
Autores principais:Madeira, Tiago de Matos Ferreira
Assunto:Computer vision 3D reconstruction Optimization Machine learning Mirror segmentation Colour correction Visual quality assessment
Ano:2025
País:Portugal
Tipo de documento:dissertação de mestrado
Tipo de acesso:acesso aberto
Instituição associada:Universidade de Aveiro
Idioma:inglês
Origem:RIA - Repositório Institucional da Universidade de Aveiro
Descrição
Resumo:Three-dimensional (3D) reconstruction is the creation of 3D models from the captured shape and appearance of real objects. It is a long term topic of investigation with numerous systems and algorithms, having gained importance in various areas, such as architecture, robotics, autonomous driving, medicine, agriculture, and archaeology. The current state of 3D reconstruction of real-world scenes is still not adequate for many applications that require photorealistic quality, including Virtual Reality (VR) and Augmented Reality (AR) experiences, and other human-centered applications. Furthermore, the best current solutions are very expensive, presenting high capital and logistical costs. This PhD aims to study and propose methods for enhancing geometry and texture in 3D reconstructions, resulting in complete virtual models with photorealistic quality. To this end we present a novel method for mirror segmentation and pose estimation, tackling artefacts introduced by reflective surfaces, a neural-based colour correction technique to improve texture consistency across multiple views, mitigating the appearance of texture seams, and introduce objective metrics to quantitatively assess the visual quality of textured 3D meshes. Extensive experiments validate the effectiveness of these contributions, demonstrating significant improvements in geometric accuracy and visual fidelity of the produced models.