Probabilistic Model for Virtual Garment Modeling - Advances in Image and Graphics Technologies - page 205

information. Chaudhuri and Kalogerakis [7,8] propose the probabilistic models for
automatically synthesizing 3D shapes using training data. These models reflect the
probabilistic graphic models that encoding semantic and geometry relationships
among shapes through segmented 3D labeled models. However, the number of 3D
garment models is limited, and it lacks diversity. The learned model may easily be
over-fitting. We train our probabilistic model on large database of garment image,
which solved the over-fitting problem.

2.3 Knowledge Acquisition from Clothing Image

The fashion coordinates system [15] uses the probabilistic topic models to learn
information about coordinates from visual features in each fashion item region(e.g.
tops, bottoms ). Our work focus on the learning of the relationship among garment
components.(e.g. collar, pocket, sleeve, and etc.) Liu, Yu[16,17] also study on
outfit combination; and in our work , we consider the combination of garment
components.

3 Overview

Our approach comprises an offline preprocessing stage in which a probabilistic graph
model is trained on a garment image database and an online interactive stage in which
the garment component parts are recommended and assembled. The overview of our
work is illustrated in Fig.1.
Offline Preprocessing. In the offline stage, the outline of processing pipeline is as the
followings:
(a) Image collection and attributes annotation. Collecting garment images from online
shopping websites. And manually annotated the garment components attributes,
such as garment style, collar style, sleeve style and pocket style, etc.
(b) Training . Learning the Bayesian graphical model which encodes the relationship
of garment attributes.
(c) 3D models cluster and annotation. The garment components are clustered by
geometry features, and then be labeled with style attributes. Thus we can use the
probabilistic graphic model learned in the training stage for components
recommendation.
Runtime Modeling. In the garment modeling stage, when the user choose a garment
body or the user select a garment style, the system update the components rank
list and the component are stitched to the existing garment using our part stich
method.