The goal of multiview 3D reconstruction is to infer geometrical structure of a scene captured by a collection of images. In our research, we concentrate on passive methods like multiview stereo, i.e. methods which do not require controlled acquisition environments and can effortlessly be applied to outdoor scenes.

In our research, we exclusively use variational approaches where, starting from a generic initialization like a box, a surface evolves towards the final reconstruction by minimizing a certain cost functional.

Below, you can see two reconstructions of well-behaved, sufficiently textured objects.

While most researchers are focused on improving the accuracy of multiview reconstruction approaches, we are also working on increasing their robustness and extending the range of applicability. As a result, our methods are able to operate in cases where most of the traditional techniques fail (e.g. shiny objects, thin small-scale structures, lack of texture).