3D Manikin Face Modeling and Super-Resolution from Range Images

In this work, a trial of modeling a manikin face
using SwissRanger SR-3000 is implemented. The
process includes acquiring data, range data
restoration, registration and surface
reconstruction. Several tests are done to evaluate
the camera s performance. Then, the noisy and low-
resolution range images are restored by MRF by
designing intensity information into the prior so
that the restored range measurements obtain the high
contrast property of the intensity information. The
range images are registered by ICP algorithm. To
improve the performance of ICP according to the
data, several variants are introduced. A new surface
reconstruction and super-resolution algorithm called
2.5D MRF is originated to combine multiple
registered surfaces. This high dimensional MRF
merges surfaces by trying to move locally smooth
patches together and keep the original values for
details. The algorithm is proved to be robust to
noise and registration errors. Finally, a face model
combined by 15 registered views via simple averaging
and super-resolution of the face combined by 3 views
via 2.5D MRF are displayed as the result.

Autorentext

Yin Yin is currently a Ph.D. candidate in electrical and computer engineering at University of Iowa. He received his B.E. in Beijing University of Technology in 2002 and M.Sc. in Technical University of Denmark in 2006. His research interest includes medical image segmentation, image processing, machine learning and computer vision.

Klappentext

In this work, a trial of modeling a manikin face using SwissRanger SR-3000 is implemented. The process includes acquiring data, range data restoration, registration and surface reconstruction. Several tests are done to evaluate the camera's performance. Then, the noisy and low-resolution range images are restored by MRF by designing intensity information into the prior so that the restored range measurements obtain the high contrast property of the intensity information. The range images are registered by ICP algorithm. To improve the performance of ICP according to the data, several variants are introduced. A new surface reconstruction and super-resolution algorithm called 2.5D MRF is originated to combine multiple registered surfaces. This high dimensional MRF merges surfaces by trying to move locally smooth patches together and keep the original values for details. The algorithm is proved to be robust to noise and registration errors. Finally, a face model combined by 15 registered views via simple averaging and super-resolution of the face combined by 3 views via 2.5D MRF are displayed as the result.