Automatic Realistic 3D Garment Generation Based on Two Images

Realistic garment model generation has a great practical significance, especially in 3D virtual try-on. However, the creation of realistic models for most 3D virtual try-on systems is labor intensive and requires strong domain knowledge. Thus, we propose a method to automatically generate realistic 3D garment model based on a front and a back images of a garment product. We classify the garment type based on the distribution characteristics of several feature points selected from silhouette points set. Besides, we propose a method to classify the component design attributes by using a classifier trained by SVM algorithm based on HOG features. Our system can create texture map as well. All of these processes can be handled automatically, which helps novice users to generate realistic 3D garments without any other manipulation after input two images. The experiments demonstrate the effectiveness and efficiency of our method on generating realistic garment models.