High-Fidelity 3D Full Head Reconstruction from Few lmages

Type Start End
National Aug 2021 Jun 2024
Responsible URL
Francesc Moreno & Xavier Giró

Description

Becent learning approaches that implicitly represent surface geometry using neural representations have shown impressive results in the problem of multiview 3D reconstruction. The effectiveness of these techniques is, however, subject to a large number (several tens) of input views of the scene and computationally demanding optimizations. This prevents them from being applicable on large volumes of data available in tech companies. ln this thesis, we will focus on the problem of full head 3D reconstruction from a few input ¡mages using implicit representations. We will research different approaches to alleviate the computat¡onal cost without hampering the accuracy of the recovered geometry. ln particular, we will take advantage of the capacity of Crisalix company to capture and annotate vast amounts of full head 3D scans, to build strong priors. These priors will then be integrated w¡thin the estimation process, to provide accurate 3D head reconstructions from very few (and eventually one) images. The outcome of this thes¡s will be integrated in the company product, an intelligent tool to assist professionals in plastic and cosmet¡c surgery, but it can also be applied to other related-applications, like in the development of avatars or online retailing apparel industry.

Collaborators