Comparing Segmentation Approaches for Learning-Aware Wireframe Generation on Human Model

Abstract

Wireframe has been proved very useful for learning human body from semantic parameters. However, the definition of the wireframe is highly dependent on the anthropological experiences of experts in previous works. Hence it is usually not easy to obtain a well-defined wireframe for a new set of human models in the available database. To overcome such difficulty, an automated wireframe generation method would be very helpful in relieving the need for manual anthropometric definition. In order to find such an automated wireframe designing method, a natural way is using automatic segmentation methods to divide the human body model into small mesh patches. Nevertheless, different segmentation approaches could have various segmented patches, thus resulting in various wireframes. How these wireframes affect human body learning performance? In this paper, we attempt to answer this research question by comparing different segmentation methods. Different wireframes are generated with the mesh segmentation methods, and then we use these wireframes as an intermediate agent to learn the relationship between the human body mesh models and the semantic parameters. We compared the reconstruction accuracy with different generated wireframe sets and summarized several meaningful design guidelines for developing an automatic wireframe-aware segmentation method for human body learning.