59-05-032 Proceeding

80 Proceedings of the Princess Maha Chakri Sirindhorn Congress 2) Experimental Results : The performance of the branded handbag recognition using each single feature and linear SVM classifier are given in Table 1. It can be seen that using the complementary feature alone will get comparable accuracy with using the original feature. The concatenation of the original feature and the complementary feature (e.g., SIFT&Complementary SIFT) performs consistently better than the single original feature, with over 2.5% improvement in accuracy. We also summarize the recognition accuracies of the 6 th layer and 7 th layer Decaf features. The 6 th layer feature leads to better results (around 6% better) than the 7 th layer feature, and is superior to traditional features. Table 1 Handbag recognition accuracies (%) using traditional and deep learning features Features Accuracy SIFT 70.02 Complementary SIFT 67.43 SIFT & Complementary SIFT 72.77 HOG 63.99 Complementary HOG 60.68 HOG & Complementary HOG 67.45 Decaf6 82.70 Decaf7 76.94 Figure 2 Mobile-based handbag recognition demo capture 3) Related Applications: A user could take a photo of a handbag using a mobile phone and find where to buy this handbag. Figure 2 demonstrates the handbag recognition flowchart. B. Shoes The development of efficient shoe tagging techniques is not only important to buyers to enrich the shopping experience, but also for the online merchants selling fashion items, where a great many new products are uploaded every day. Here, the annotation of shoes corresponds to