59-05-032 Proceeding

77 Proceedings of the Princess Maha Chakri Sirindhorn Congress VISUAL PRODUCT SEARCH FOR FASHION DOMAIN: FROM HAND-CRAFTED FEATURES TO DEEP LEARNING Huijing Zhan 1 , Yan Wang 1 , Abrar H. Abdulnabi 1 , Sheng Li 2 , Dennis Sng, Alex C. Kot and Simon See 3 Abstract : Due to the huge profits from e-commerce especially in the fashion domain, a large body of research has been conducted to enable computers to intelligently perceive and analyze the objects visually, which involves multi-disciplines like computer science and artificial intelligence. Also, different categories of fashion products have been explored like clothing, handbag and shoes. For efficient description of those objects, the selection of an appropriate feature representation scheme is of great importance. In recent years, due to the unmatched performance of deep learning methods, the trends move from using the traditional hand-crafted features to automatically exploring the appropriate feature for representation by analyzing the big data. In this paper, we use inter-disciplinary knowledge fromcomputer science, engineering and systemarchitecture to study fashion-related object search. Existing techniques are reviewed and newmethods are introduced. Preliminary experimental results on product search are presented using both traditional and deep learning approaches. Keywords : visual object search, e-commerce, deep learning, computer vision, fashion search I. Introduction Fashion profits have occupied a large portion of the entire market. It is reported that in the clothingmarket alone, total retail sales have reached up to about $980 billion in 2012 [1]. Due to its promising market size, research in the fashion field has been receiving more and more attention, especially the intelligent fashion analysis. It enables computers to analyze the fashion product more smartly using vision-based approaches, which incorporate multi-discipline knowledge from computer science, engineering. Currently, researchers in this field have been developing algorithms addressing different types of fashion products, like handbags [6], shoes [9], clothing [2], [3] and the list will go on. To efficiently employ vision based methods on objects, firstly it is essential to develop an appropriate feature representation scheme. Conventionally, hand-crafted features like SIFT, HOG, GIST are popular choice for their powerful ability to capture the salient semantics of objects. However, in recent years, deep learning based methods [10], which model high-level abstractions of objects [11], have become increasingly popular, outperforming state-of-the-art traditional methods in handling various vision tasks and different objects as noted in [13], [14]. What is more, it makes good use of multidisciplinary knowledge like neural networks in computer science, 1 Rapid-Rich Object Search (ROSE) Lab, Nanyang Technological University, Singapore 2 School of EEE, Nanyang Technological University, Singapore 3 NVIDIA Corporation, Santa Clara, USA