59-05-032 Proceeding

81 Proceedings of the Princess Maha Chakri Sirindhorn Congress providing labels for a shoe, for example, given a pump shoe input image, the computer is capable of automatically generating descriptions like `high heel’ and ‘has back cover’. Few works focus on addressing the shoe annotation tasks, which describe high-level concepts of shoes. We build a novel a multi-view shoe dataset, on which traditional and deep learning features are tested to evaluate the attribute prediction result. 1) Pump ShoeDataset :The images of our pump shoe dataset are collected fromAmazon. comwith clean background. In total, our pump shoe dataset consists of 7500 shoe images inmultiple viewpoints. For each shoe, the ground truth annotation contains seven binary part-aware attributes defined based on four different shoe parts, as shown in Table 2. The ground truth annotation of each attribute is collected manually. Figure 3 displays the multiple viewpoints settings for a shoe. Figure 3 Multiple-view display settings for a shoe Table 2 Shoe parts and their part-awareness attributes Shoe Parts Related Attributes head Closed toe, pointy body Side-covered, bounds Back Back-covered heel High thin heel, wedge heel 2) Experimental Results : Table 3 shows the results of shoe attribute annotation using the traditional features SIFT and HOG in comparison with deep learning features Decaf7 and Decaf8 activated from the full-connected layers fc7 and fc8 of the deep learning multi-layer networks. The performance is evaluated using the Mean Average Precision metrics. In the Table, we use the bold numbers to highlight the best-performing feature. From the results, we could find that even though traditional features performs well on annotation tasks, however, employing deep learning features improved attribute prediction significantly. Also, among all the attributes defined, Decaf8 achieves slightly higher prediction accuracy thanDecaf7. It is reasonable for the last full-connected layer to provide more powerful and hierarchical representations for shoe shapes.