Deep Part-Based Image Feature for Clothing Retrieval

Abstract

In this paper, we propose a straightforward way to extract part-based features only with the supervision of part-based attributes. As we know, regions can be highlighted by labels through weakly-supervised segmentation algorithms, and deep features can be extracted from CNN convolutional layers. We develop a new approach to combine them, leading to simpler procedure with only one CNN forward pass and better interpretation. We apply this method to our database of over 100,000 clothing images, and achieve comparable results to the state of the art. Moreover, the part-based features support functionalities of tuning weights among the parts, and substituting visual part features from other clothes. Because of its simplicity, the method is promising to be transferred to other image retrieval domains.