Revisiting Skeletons from Natural Images

In the last two decades there have been several works promoting shape fields that implicitly encode local convexity/concavity properties of the shape boundary. These shape fields are formulated either as solutions to Poisson type PDEs or via heuristic approximations to them. The v-field of Tari-Shah-Pien, can be computed directly from a real image; thus, suggests a mechanism to bridge low level visual processing and high level shape computations. We revisit Tari, Shah and Pien's v-field approach and extend its application to complex images with texture. We relate v-field value at a skeleton point to the distance of the point from a putative shape boundary, and use this relation to extract semantic image patches. At the end of the chapter, we experimentally compare the medial locus computed from the new v-field to that of Kimia et al.