Probabilistic Model for Virtual Garment Modeling - Advances in Image and Graphics Technologies - page 207

Fig. 2. Representing distributions of clothing components with a Bayesian network. Top: a
Bayesian network for clothing components, trained by the labeled clothing image database.
Bottom: a table showing the random valuables node and descriptions.

4.1 Probabilistic Graphic Model

Our probabilistic graphic model encodes the joint distribution on garment style and
components. The purpose of the model is to recommend garment components
compatible to the existing models. The hierarchical graphic model is showed in Fig.2.
It contains G .as the root represents the garment style, and contains random valuables
such as collar, sleeve, garment body, pocket, belt and button. which represent
,, ,,,

, 0 represents none, and nonzero
represents the style attributes of the component. For

CSM PBT respectively. Z + CSM P ,, , ∈∪ {0}

0 means none and 1
means exist. As the belt and button style is not as complex as others, we reduce the
model complexity just by considering their existence. The graphic model may contain
lateral edges between the nodes representing the strong dependency between
components, which can be learned from the structure learning.
The conditional distribution of discrete random variable of

BT , ∈ , }

CSM PBT can be
represented as conditional distribution table(CPT). Considering a discrete variable
Y with a single parent discrete variable X , each assignment y to Y and x to X the
CPT at Y contains the entry:

,, ,,,

PY ( = y X | = = x ) q

(1) yx |

The values Qq =

{}

is the parameters of the CPT. For the root node G , the

yx

parameters comprises .

We define the joint distribution as

PG ( == g ) q

g

PU , when a component is selected, the node
is set as observed variable and the unobserved variables are query variables. For
example, when the stand-collar is selected, C became observed and is assigned to 1, or
if the garment style is set as T-shirt, the node G became observed.
During the inference process, when given the observed variables, we compute the
probability of query variables. Assuming given O , which contains the subset of

()