Over the decades and centuries, many statues displayed in cultural monuments have been damaged. This is for example the case of the Virgin with the Child in Lausanne’s Cathedral, whose head has been missing for centuries. With the recent impressive progress of deep machine learning models, one can thus hope to generate this missing part, so as to better understand our cultural heritage and display it to the public. The goal of this project therefore is to study the use of the recent developments in Diffusion Model, Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (3DGS) to digitally reconstruct in 3D the head of the Virgin with the Child statue in Lausanne’s Cathedral. As no visual information about this missing head is available, the student will investigate ways to learn prior 3D and appearance models from similar statues in other Cathedrals and Churches.
Madonna and Child, Lausanne Cathedral
From left to right:
1) Virgin with the Child from the Sainte-Chapelle, before 1279, in ivory.
2) Enthroned Virgin with the Child, ivory statuette, H. 24 cm, W. 10 cm, crafted around 1240-1250 in Paris, preserved in the National Museum of the Middle Ages section of the Cluny Museum in Paris.
3) Virgin with the Child, also known as Virgin in Majesty, second half of the 13th century. Walnut wood applique sculpture with hollowed back, featuring minimal traces of polychrome in the recesses. Restoration includes the Virgin’s right forearm, a later addition of the cross on the globe, and missing polychrome. Visible damages and losses are noted.
Prerequisites:
- Python
- Pytorch
- Knowledge of deep learning and computational graphics
Contact:
References:
[1] CLAY: A Controllable Large-scale Generative Model for Creating High-quality 3D Assets, Siggraph 2024
[2] 3D Gaussian Splatting for Real-Time Radiance Field Rendering, Siggraph 2023
[3] 2D Gaussian Splatting for Geometrically Accurate Radiance Fields, Siggraph 2024
[4] SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Qualitiy Mesh Rendering, CVPR2024
[3] Neuralangelo: High-Fidelity Neural Surface Reconstruction, CVPR2023