Specular surface reconstruction from images is a relatively little explored topic, due to the rareness of such objects and the inherent complexity in the induced geometry compared to that of matte shading. This is due to the fact that the apparent texture of the surface is dependant on the viewpoint, or in other words, that the light path from a feature point to a camera pixel is not a straight line. As such, these surfaces are ignored or considered as noise in most 3D reconstruction techniques, whereas we show that the constraints they can provide can be used to compute accurate geometric information on position and orientation. This thesis presents two methods for obtaining position and orientation of points of a specular surface from the reflection of known environning points, without making the usual smoothness or continuity assumptions of nearly all prior methods. The first of these methods is an extension to the space carving framework, which computes voxels belonging to a specular surface by relying on geometric consistency measures rather than photometric ones, which in this case do not make sense. The second method is based on triangulation, and requires a fixed camera viewing the reflection of at least 2 environning points for each surface point to be reconstructed. Finally, we present methods for estimation of the pose of calibration objects when they are not in direct view of a camera, through the reflection by specular objects. The first of these methods assumes that the calibration object is seen through 3 or more mirrors known to be planar, while the second presents a theoretical constraint for the obtention of the pose of such an object placed at two locations, when seen through the reflection of an arbitrary surface.
