Visual Prototyping of Knitwear

Abstract

In this thesis, we address the problem of prediction and editing of the virtual appearance of knitted cloth. Among the most accurate approaches to model fabrics and other complex materials is to represent them in terms of dense Bidirectional Texture Functions (BTFs) that describe the spatially varying material appearance under different viewing and lighting conditions parametrized on a proxy geometry. This requires taking an exhaustive set of photographs under different light and view directions to accurately reproduce the appearance of an existing material, and thus comes at a high cost regarding storage requirements, motivating the research of convenient compression methods. The current state-of-the-art neural approach did not fully exploit the capabilities of the proposed architecture, as the compression ratio did not depend on the complexity of the material. Therefore, in the first project presented in this thesis, we developed an approach to even further compress compact BTFs, that were compressed based on state-of-the-art neural compression, and let the compression ratio depend on the complexity of the material. <br> Besides high memory requirements, the usage of BTFs does not allow for easy editing operations. Moreover, BTFs are not suitable for macro-scale geometry representation. Both these limitations pose a problem in the context of visual prototyping of knitwear. There are three different levels of knitted cloth design that are important for our task. First, there is the choice of the yarn, which in turn consists of fibers of the same type or a mix of different fiber types. Then, the selected yarn is knitted into a chosen pattern producing the actual knitted cloth. This thesis aims to enable editing operations on all three scales: patterns, yarns and fibers. To this end, in two projects that constitute the second part of this thesis, we develop algorithms and models for macroscale representation of knitting patterns, mesoscale modeling of yarns, and visualization and easy editing of knitting yarns, fibers and patterns that we automatically induce from images of real yarns or knitted pieces, respectively.