Probabilistic Model for Virtual Garment Modeling - Advances in Image and Graphics Technologies - page 206

Fig. 1 system overview

4 Bayesian Networks for Clothing Components Relationships

Clothing Image Dataset and Attributes Annotation. The garment images are
crawled from the online shopping websites by using the keywords such as “T-shirt”,
“coat”, and “suit” , etc. The images are classified by the keywords of style. The
garment style is affected by the garment components attributes, for example, a T-shirt
always has a short sleeve and has no button, a suit always has a flat collar and has
buttons but do not have belt, etc.
We manually labeled the garment components. We define the garment component
parts attributes and labeled about 500 clothing images for Bayesian network training.
The attributes of collar, pocket, button, belt of each image are labeled.
Our system trains the Bayesian network for garment components for five garment
styles: T-shirt, shirt, skirt, coat and suit. Figure 2 shows a part of the Bayesian
network for garment components. The nodes of garment components have different
attributes we have defined.
The node of Bayesian network represents different garment component parts and
each node state represents the style attribute of garment component. For example, the
node collar style has attributes stand-collar, fold-collar and flat-collar. And each
component node has a state, but none of which indicates that the garment does not
contain that component.