Abstract
The focus of this paper is on retrieval of fashion images after manipulating attributes of the query images. This task is particularly useful in search scenarios where the user is interested in small variations of an image, i.e., replacing the mandarin collar with a button-down. Keeping the desired attributes of the query image while manipulating its other attributes is a challenging problem which is accomplished by our proposed network called FashionSearchNet. FashionSearchNet is able to learn attribute specific representations by leveraging on weakly-supervised localization. The localization module is used to ignore the unrelated features of attributes in the feature map, thus improve the similarity learning. Experiments conducted on two recent fashion datasets show that FashionSearchNet outperforms the other state-of-the-art fashion search techniques.