Probabilistic Model for Virtual Garment Modeling - Advances in Image and Graphics Technologies - page 204

In previous work, Chaudhuri, Kalogerakis [7,8] analyzed a large set of segmented
3D models and used a probabilistic graphic model to learn their semantic and
geometry relations for the exploration of a smarter design. As this work relies on the
existed 3D segmented models, we bypass the difficult work by learning from a large
set of garment images.
In this paper, we present a probabilistic graphic model called Bayesian network
[9,10] to automatic recommend garment components to the users during 3D garment
modeling process. A probabilistic graphic model is well-suited to encoding the
relationship between the random variables such as the garment style, sleeve style,
collar style, pocket style, existence of belt and button, etc.
Our main contribution is that we propose a part assembly method for garment
design, and we lean a probabilistic graphic model from images. The system
automatically recommends garment components and we also propose an improved
mean value coordinates [11] method for part stich.
We demonstrate the effectiveness of our model using the part assembly garment
modeling tool that we have developed. Experiment shows that the probabilistic model
produces more relevant recommendations than a static presentation of components
could.

2 Related Work

Part assembly modeling remains a popular research topic, but we built our work on
some of the ideas and algorithms. In this section, we present a few representative
papers relevant to our work.

2.1 Part Assembly Modeling

As the model collections grow, the assembly-based modeling provides a quick way to
create new models by reusing the existing models. The modeling-by-example system
relies on shaped-based or text-based search to retrieve component parts [5].
[12,13]propose various sketch-based user interfaces: In these methods, the user must
be very clear about the specific component. This is not appropriate for 3D garment
modeling when the user has no design expertise. Kreavoy [6] described a method in
which the user can interchange parts between a set of compatible shapes. This method
requires the shapes that share the same number of components. Chaudhuri [7]
proposed a data-driven technique that can recommend components to augment a
given shape. This method just considers the geometry feature but doesn't take into
account the semantics of components.

2.2 Probabilistic Framework

Fisher [14] describes a probabilistic model for 3D scene modeling. The user drew a
boundary box, the model extracts a model from database using the context