59-05-032 Proceeding

82 Proceedings of the Princess Maha Chakri Sirindhorn Congress Table 3 Attribute prediction accuracy (%) using traditional and deep learning features Attribute Type HOG SIFT Decaf 7 Decaf 8 Closed toe 83.91 80.36 86.23 87.13 Pointy 91.18 90.61 90.73 91.20 Bounds 87.87 87.17 88.90 89.22 High-heel 88.50 88.32 90.52 90.37 Wedge-heel 94.95 94.37 95.83 94.78 Side-covered 89.98 88.90 94.10 94.48 Back-covered 95.34 94.93 96.64 97.97 3) Related Applications : The proposed system can be applied to tag the shoe images, which saves the merchants’ effort in creating shoe descriptions. C. Clothing In on-line fashion shopping, clothes such as dresses are top selling products among feminine marketplace. This leads to the need to develop vision programs that are able to predict which dress items are more attractive, thus having the potential to be placed on the top of the website pages. 1) Experimental Result: To tackle the problem of extracting robust attribute feature representations, we adapt deep feature learningmethods like convolutional neural networks (CNN) [10]. Specifically, we adapt deep CNN to learn binary semantic attributes, where each CNN will predict one binary attribute, hence eachCNNwill generate attribute-specific feature representations. We train these CNNmodels on the Clothing attribute dataset [5] including clothed people mostly pedestrians on the street with cluttered background as shown in Figure 4. In Table 4, we show the mean average accuracy prediction of attributes to compare the performance of some baseline Figure 4 Example of clothed people images from the Clothing Attribute database [5]