SMARTPHONEHEADSCANNER-Intelligent Solution for 3D Head Reconstruction on a Smartphone using Neural Radiance Fields

• Autor(a) principal: Correia, Helena Araújo
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/11110/2715
• topic: 3D reconstruction of heads; Multi-View; Smartphone; 3D Deep Learning; Neural radiance fields

This study addresses the topic of multi-view 3D reconstruction applied to the area of preventive medicine. From this perspective, the use of 3D scanners for medical procedures, namely for the prevention, diagnosis, and therapy of deformities in the human body, is restricted to medical facilities with access to these resources and to specialized clinicians able to manipulate these same devices. Thus, the existence of standard devices capable of performing the same functions as a 3D scanner is crucial for parents to monitor the evolution of head growth at home to prevent deformities such as positional plagiocephaly. This dissertation takes a step towards this goal, presenting an intelligent system of three-dimensional multi-view reconstruction of heads in real-time using a smartphone for data acquisition and visualization of the 3D model. For this purpose, a mobile application was developed to acquire videos and the intrinsic camera parameters used for the three-dimensional reconstruction process. The videos are captured from circular movements from different points of view to cover an area of 360º around the head. Furthermore, the application is responsible for exposing the user to the 3D model obtained from the reconstruction process intuitively and interactively. During this project, the evolution of 3D reconstruction in recent years was investigated, performing a study in the literature to identify innovative methods that best fit the problem of multi-view 3D reconstruction. A method based on neural radiance fields was selected due to its fascinating results. A second study was conducted to evaluate the method's robustness, specifically for 3D head reconstruction. To this end, several experiments were performed to determine the ideal parameters (number of input images and epochs) to obtain precision and detailed 3D models. In conclusion, the results express the potentiality of applying neural radiance fields for the three-dimensional reconstruction of heads, with dolls' heads and real heads as case studies. With the results obtained, it is possible to perform an objective evaluation regarding the head shape, preventing future deformities through an intervention of the relatives in the preferential position of the newborn. It is worth mentioning that it will be crucial to carry out prospective studies to evaluate the system's performance in newborns' heads. Overall, the presented dissertation successfully expands the topic of 3D reconstruction to the area of preventive medicine.