3.2. RELATED WORK 13
Top
Pre-trained
CNNs
C: Women’s Fashion ->
Tops->Blouses.
T: Off-Shoulder Blouse.
C: Women’s Fashion ->
Skirts->Mini Skirts.
T: Plaid Ruffled Mini Skirt.
Bag of Style
Phrases
Pre-trained
CNNs
Bag of Style
Phrases
Bottom
Context ImageContext Image
Feature Extraction Latent Compatibility Space
Bayesian Personalized Ranking
Figure 3.3: Illustration of the proposed scheme. We employ a dual autoencoder network to learn
the latent compatibility space, where we jointly model the coherent relationship between visual
and contextual modalities and the implicit preference among items via the Bayesian Personalized
Ranking. C: category, T: title. “ >” indicates the category hierarchy.
3.2 RELATED WORK
3.2.1 FASHION ANALYSIS
e fashion domain recently has been attracting increasing attention from both the computer
vision and multimedia research communities. Existing efforts mainly focus on clothing re-
trieval [77], clothing recommendation [42, 76], and fashionability prediction [67, 104]. For
example, Liu et al. [76] proposed a latent Support Vector Machine (SVM) [21] model for an
occasion-oriented outfit and item recommendation, where the dataset of wild street photos was
created by human annotation. Iwata et al. [48] proposed a topic model to recommend tops for
bottoms with a small dataset collected from magazines. Due to the infeasibility of a human-
annotated dataset, several pioneering studies have resorted to other sources, where rich data can
be harvested automatically. For example, Hu et al. [43] studied the problem of personalized
whole outfit recommendation over a dataset collected from Polyvore. McAuley et al. [86] pre-
sented a general framework to model human visual preference for a pair of objects based on the
Amazon co-purchase dataset. ey extracted visual features with convolutional neural networks
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.