To reinforce application of AI to the fashion industry, Hong Kong Polytechnic University (PolyU) has developed the world's first FashionAI Dataset in collaboration with e-commerce giant Alibaba to better address the needs of fashion retailers and consumers, according to Calvin Wong, a professor at PolyU's Institute of Textiles and Clothing.
At a recent Taiwan-Hong Kong economic cooperation forum in Taipei, Wong said that through integrating fashion knowledge and machine learning formulation, the Dataset will enable machines to better understand fashion, ushering in a new horizon to the fashion retail industry through the application of AI.
Dataset can solve two fundamental problems of deep learning algorithms: detection of apparel key points and recognition of apparel attributes. Key points to be detected include neckline, cuff and waistline; and attributes to be recognized cover sleeve length, collar type and skirt style. Both build the foundation for machine learning in understanding fashion images, according to Wong.
He said Dataset can help improve the accuracy of online fashion image searching, create innovative buying experience and facilitate customization of online shopping platforms.
E-commerce is now a major market for the fashion industry with many fashion brands scoring over 50% of revenues from online sales, Wang said, adding that some cosmetics brands have applied AI-based facial recognition technology to offer virtualized trial makeup services.
Wong indicated that e-commerce operators, besides harnessing shopping habits and preferences, are also seeking to better identify customers' body shapes and sizes to enhance promotions and reduce return rates for sold apparels.
To meet the demand, Wong disclosed, PolyU has developed an intelligent 3D human modeling system by integrating cutting-edge computer graphics and vision technology. The system allows a customized model in arbitrary dynamic poses to be created automatically within 5-10 seconds, able to detect body parts in tight fitting clothing from front-view and side-view 2D images and then construct them into 3D models.