The January 2022 JCISE Spotlight presentation by Tsz Ho Kwok (Concordia University) is now available for viewing on YouTube. In this free presentation Dr. Kwok discusses his recent JCISE paper (December 2021 issue), “Segmentation-Based Wireframe Generation for Parametric Modeling of Human Body Shapes“.
Abstract: Deep neural networks can learn complex relationships, but they depend heavily on the setup of hidden layers. For parametric modeling of human body shapes, which needs to learn the relationship between the human body and semantic parameters, we have proven that using wireframes as an intermediate layer can improve the learning performance. However, the definition of the wireframe must have anthropological meaning and depends highly on experts’ experience. Hence, it is usually difficult to get a well-defined wireframe for a new set of shapes. An automated wireframe generation method would help relieve the need for the manual anthropometric definition to overcome such difficulty. One way is to apply segmentation to divide the models into small mesh patches. Nevertheless, various segmentation approaches could have distinct segmented patches, thus resulting in a variety of wireframes. How do these different wireframes affect learning performance? This research attempts to answer this question by defining several critical quantitative estimators to evaluate the learning performance. To find how such estimators influence wireframe-assisted learning accuracy, we conduct experiments by comparing different segmentation methods on human body shapes. We summarized several design guidelines for the development of automatic wireframe-aware segmentation methods for human body learning with such verification.