• 3D face reconstruction
• 3D reconstruction

Nowadays, 3D reconstruction from images has played an important role in computer vision with many improvements in both quality and performance. One of its main uses is the generation of 3D models of objects that are difficult to model.
This work mainly focuses on creating 3D models of human heads from 2D images using neural representations. However, these techniques have a significant limitation as their effectiveness is strictly dependent on the availability of a large number (several tens) of input views of the scene, and involve computationally expensive operations.
In this work, the specific problem of full 3D head reconstruction is addressed by adapting coordinate-based representations with a probabilistic prior that allows for faster convergence and better generalization when using few input images. The reconstruction is done in 2 steps, first we learn a 3D model of the head shape from different point clouds using implicit representations. After that, the learned prior is used to initialize and constrain the geometry of the reconstruction.
By doing so, we obtain high-fidelity reconstructions of the head, including the hair and shoulders, and with a high level of detail that exceeds state-of-the-art methods.