Correspondência de Pontos em Formas 3D Baseada em Aprendizagem Profunda Multivisão

• Ano de defesa: 2022
• Autor(a) principal: Alexandre Soares da Silva
• Orientador(a): Paulo Aristarco Pagliosa
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Fundação Universidade Federal de Mato Grosso do Sul
• País: Brasil
• Palavras-chave em Português: Processamento geométrico; Correspondência de pontos; Formas 3D; Aprendizagem profunda multivisão
• Link de acesso: https://repositorio.ufms.br/handle/123456789/5632

In the field of geometric processing, several techniques proposed in the literature require the establishment of correspondence points between two or more surfaces, that is, given a point on a source surface, it is necessary to associate which point on a target surface corresponds to the given point. Applications include surface reconstruction, cross-parameterization, pose transfer, texture or animation transfer, shape recognition, and search, among others. Defining a mapping function between two shapes, even for a discrete number of characteristic points, does not always involve only geometric or structural relationships, but also semantic relationships. Since such a mapping cannot generally be directly expressed by purely axiomatic approaches, in various geometric processing methods, the indication of an initial set of correspondence points is manually performed, through processes that can be laborious and error-prone. In fact, discovering semantic relationships between any shapes without any user interaction is still considered an open problem. Machine learning models, especially deep learning, have evolved due to their ability to use large datasets to estimate the solution to problems in various areas of knowledge, including geometric processing. This work presents a method that uses deep multiview learning as part of the processing responsible for finding automatically, that is, without direct user intervention, correspondence points between 3D shape surfaces represented by triangle meshes. The method is divided into two components: training and correspondence. The former is a multiview training that learns, with the aid of a CNN, to detect feature points in 2D images derived from triangle meshes of the training set. The latter uses the results of the training to infer semantic correspondences with feature points (vertices) in 3D shapes. The discovery of these points does not require new training or human interaction during the correspondence pipeline.