Abstract
This thesis presents a new shape representation called Unraveled Skeleton. It is designed to be robust to a large number of transformations, especially articulated movement. Based on this new representation, its general properties are studied and its application domain established. The representation is based on the skeleton or medial axis of a shape. It uses the very robust Voronoi skeletonization and "walks" around the resulting skeleton tree. At every point the minimum distance to the boundary is saved to a vector, resulting in a list of distance measures, a shape signature. This Unraveled Skeleton vector can then be used for further processing like normalization. This thesis explores two possible directions to use the Unraveled Skeletons: Shape analysis and pose independent coordinate systems. In shape analysis the Unraveled Skeleton vector can be used for articulation independent recognition. For this purpose a number of different normalization and optimization techniques are introduced. It is also possible to use parts of the vector to match parts of objects. The pose independent coordinate system assigns every point inside the shape and on its boundary a unique coordinate independent of pose or articulated movement. Using the closest points on the skeleton and its distance as provided by the Unraveled Skeleton one can address convex patches. Given a third component to chose a point from this patch a coordinate with three components is introduced. This paper states three applications to use the Unraveled Skeleton coordinates: Shape Blending, Super-resolution and Segmentation refinement. It also explores possibilities of non-centered skeletons in regard to Unraveled Skeleton coordinates. All results are evaluated on multiple datasets.
Reference
Langer, M. (2019). Unraveled skeletons [Diploma Thesis, Technische Universität Wien]. reposiTUm. https://doi.org/10.34726/hss.2019.65422