Project B05

B05 | Efficient Large Scale Variational 3D Reconstruction

Prof. Bastian Goldlücke, University of Konstanz
Email | Website

Prof. Andrés Bruhn, University of Stuttgart
Email | Website

Ole Johannsen, University of Konstanz – Email | Website

About this project
Results

The central goal of the project is to research and develop high-performance variational methods for large scale 3D reconstruction problems, which are general and accurate while meeting computation time constraints imposed by visual computing applications. Key abilities will be that

a variety of possible sources of 3D information can be integrated,
accuracy can be locally improved at the cost of run-time, and
for a given allowed total run-time, global accuracy is maximized according to application-specific metrics.

Research Questions

What is the best geometric representation to implement locally adaptive geometry optimization?

Which variational models efficiently allow local refinement of the scene structure in a mathematically consistent manner?

Which models in shape reconstruction best translate to the adaptive shape representation?

How can the local refinement of surface properties (e.g., texture) be modeled and optimized—ideally together with the geometry?

What are natural, ideally convex priors on adaptive grids?

What are optimal local accuracy measures for different applications?

How can an ideal trade-off between accuracy and run-time be determined?

Results on synthetic data with semi-reflective statue of Warrior. Top images depict center views of corresponding input light fields. Bottom resulting geometry.

Fig. 1: Results on synthetic data with semi-reflective statue of Warrior. Top images depict center views of corresponding input light fields. Bottom resulting geometry.