PhD defence by Robin Bruneau

Title

3D Reconstruction with Refraction

Abstract

Amber is a fossilised resin from ancient conifers. It can be found in many places, and if you are careful, you can discover relics of the past embedded inside. Whether plants or insects, these fossils are millions of years old. At a time when the evolution of the living world around us is being studied, entomologists would like to be able to study these insects in detail so that they can be linked to the insects of today. While one basic method would be to observe them under a microscope, another would be to provide a 3D model of these insects so that they can be observed more easily, along with the ability to take measurements of them. This thesis investigates methods for the 3D reconstruction of objects immersed in refractive media, with the idea that in future work, the complex case of amber will be addressed. The research integrates geometric and photometric methods, as well as neural inverse rendering techniques to develop robust and precise 3D reconstruction pipelines. The first focus is a method based on multi-view, which provides a triangular mesh of a polyhedron using corner detection and ellipsoid trajectory estimation. With the knowledge of the shape of the medium, work was then carried out on adapting multi-view stereo to refraction. At the same time, the adaptation of photometric stereo in the case of a refractive planar interface between the camera and the object was also studied. Finally, following the work on the neural methods for 3D reconstruction, we proposed a pipeline for 3D reconstruction with high details using multi-view multi-illumination images. The latest work, which is not yet fully conclusive, has focused on obtaining refractive normal maps and on the multi-view integration of these normal maps within the framework of a polyhedral refractive interface.

Supervisors

Principal Supervisors: François Lauze & Jean-Denis Durou
Co-Supervisors: Yvain Quéau & Kim Steenstrup Pedersen

Assessment Committee

Professor Sune Darkner, DIKU
Research Director Peter Sturm, Institut de Recharche en Informatique et Automatique (INRIA) Grenoble, France
Associate Professor Tomas Pajdla, CIIRC – The Czech Institute of Informatics, Robotics, and Cybernetics, CTU in Prague

For an electronic copy of the thesis, please visit the PhD Programme page.