Probabilistic Model for Virtual Garment Modeling - Advances in Image and Graphics Technologies - page 208

{,, , , ,} CSM PBT ,

q UU

∈ , q is the query variable, and we define the score as the

marginal probability:

score U () ( = P U = q | O )

(2)

q q

Learning
The input of the offline learning process is a vector set D representing the labeled
clothing image dataset.

 . We have K
training images in total. The output of the learning process is a directed graph and the
conditional probabilistic table(CPT) for each node.
The best structure G is the one that has highest probability when given training
data D (Kollar and Friedman 2009). By Bayes' rule, the probability is as follows:

DCSMPBT = {,, ,, ,}

i

, i = {1, 2 , K }

i i i i i

PD GPG (| )()

PG D (| ) = (3)

PD ()

Cooper [18] assumes a uniform prior distribution

pG over all possible ()

structures. Thus maximizing PG D is equal to maximizing (| ) PD G . We define (| )

as the prior distribution over structure G ,then the marginal likelihood can be
expressed as:

ʸ



PD G (| ) = PD G (| ,)(| ʸʸ ʸ P Gd ) (4)

ʸ

We adopt the K2 algorithm to the structure learning The K2 algorithm[18] is a greedy
search algorithm that works in the following ways. Initially each node has no parents.
It then incrementally adds the parent whose addition increases the score of the
resulting structure most. When the addition of no single parent can increase the score,
it stops adding parents to the node. After performing the K2 algorithm, the structure
and the CPT are learned.

4.2 Clothing Components Recommendation

3D Model Clustering: the object of our system is applying the implicit relationships
among the garment components learned from images to 3D detailed garment
modeling. Firstly, we need to establish the relationship between the images and 3D
models. During the training stage, we labeled the style attributes of each component,
for example, the collar style is labeled as stand-collar, fold-collar and flat-collar.
Thus, we also labeled the style attributes for each 3D garment component. To
annotate the 3D garment component models, we use K-Means classifier to cluster the
models. For 3D model feature extraction, we adopt the visual based method [19] , and
in his method, each model is rotated twelve times and be projected in 10 direction.
Although it has high recall and precisian, it costs much more time. In our system, we
simplify this method by aligning the model first, we use PCA to find the three main
component axis and get the project view from the three direction; And then we extract
the shape moment feature for clustering.