Fast Fashion Guided Clothing Image Retrieval: Delving Deeper into What Feature Makes Fashion

Clothing fashion represents human’s aesthetic appreciation towards their outfits and reflects the development status of society, humanitarian and economics. Modelling fashion via machine is extremely difficult due to the fact that fashion is too abstract to be efficiently described by machine. In this paper, we delve into two fashion related problems: what type of image feature best describes fashion and how can we fast retrieve the fashionably similar images with any given query fashion image. To address these two problems, we first conduct extensive experiments on various image features, ranging from traditional low-level hand-crafted features , mid-level style aware features to current high-level powerful deep learning based features, to find the feature best describes clothing fashion. To test each candidate feature’s performance, we further design a fast fashion guided clothing image retrieval framework by efficiently converting float formatted features into binary codes, with which we can achieve much faster image retrieval without much accuracy reduction. Finally, we validate our proposed framework on two publicly available datasets. Experimental results on both intra-domain and crossdomain fashion clothing image retrieval show that deep learning based image features with explicit fashion prior knowledge guidance best describe fashion, and feature binarization scheme also achieves comparable results in terms of various fashion clothing image retrieval tasks.