| Abstract |
This thesis considers texture reconstruction for scanned 3D models.
Given a geometric model and several photographs of the object, the
texture is reconstructed in two steps: firstly, the images are registered
(aligned) to the model, and, secondly, the texture is constructed from
images. We split the first problem into initial registration, followed by
optimization of the registration, and focus on the optimization part.
We propose a framework which registers the images jointly, exploiting
the model-image and image-image relations, using the mutual information criterion. The optimization uses a stochastic gradient-based
algorithm, and its time complexity does not depend on the resolution of the data. We applied the framework to several models, and
we achieved accuracy in the order of one pixel. We propose a novel
evaluation method using epipolar geometry, and analyze three measures that allow comparison of texture registration with camera calibration data (weak calibration). The method is intended to detect
biases of the texture registration. The proposed measures are well
known in computer vision, and we investigated new aspects about
them. We compared our texture registration algorithm with a state
of the art camera calibration algorithm, and confirmed the high accuracy of our method. Finally, we developed a multi-band blending
algorithm, based on the partition of unity over a mesh, to build a
seamless texture. |