Quality guided handbag segmentation

In this paper, we address the problem of handbag segmentation, which is a challenging while important pre-processing for fashion related applications such as handbag tagging and search. Inaccurate segmentation will easily lead to other descriptions of color and shape of the handbag. We first design and extract a set of features for measuring the quality of the handbag segmentation based on some prior knowledge of handbag images. The quality of the handbag segmentation is then measured based on the weighted combination of these features. Guided by such quality measurement, we propose to segment the handbag image by a bottom-up super-pixel fusion. We conduct the experiment on a newly built handbag dataset as well as an existing branded handbag dataset. The results show that our segmentation algorithm performs favorably for handbags. The performance of handbag tagging and recognition is shown to be improved by incorporating such algorithm as pre-processing.