Augmented reality mixes computer generated graphics with real imaging using computer
vision techniques. However, nowadays, augmented reality is still a very young field of
research, and its applications usually involve predefi ned tags. This thesis has been
directed to use computer vision and arti cial intelligence techniques to explore the
viability of using natural landmarks as key points for computer graphics reference.
Moreover, there are many techniques that infer 3D scenes from images like stereo-vision,
structures from motion, depth images, shape from shading, etc. The aim of this work
is to find a way of doing this from one single shot image. Finally, new virtual elements
will be integrated on the final scene using contextual colors.
The followed methodology has been to automatically segment an image in small planar
surfaces using di fferent granularities of small regions. Each region is assumed to likely
lie on only one planar surface, and thus it the 3D face that it came from can be inferred.
The normal vector of the planes corresponding to the 3D faces are approximated along a
discrete set of orientations. In addition, some regions do not have a regular orientation
and thus, they are assumed as a texturized or porous region.
Inferring the fi nal 3D orientation and location from the set of labelled regions is a
non-trivial task. This work proposes a method based on the coherent topology of the
neighborhood. The 3D position of each point of a region is found and a 3D scenario
can be obtained. After that, the regions of the original images are textured in the 3D
reconstructed faces.
Finally, a color transfer approach is used to integrate new 3D objects inside the final
scene.
