Hierarchical Category Detector for Clothing Recognition from Visual Data

Abstract

Clothing detection is an important step for retrieving similar clothing items, organizing fashion photos, artificial intelligence powered shopping assistants and automatic labeling of large catalogues. Training a deep learning based clothing detector requires pre-defined categories (dress, pants etc) and a high volume of annotated image data for each category. However, fashion evolves and new categories are constantly introduced in the marketplace. For example, consider the case of jeggings which is a combination of jeans and leggings. Detection of this new category will require adding annotated data specific to jegging class and subsequently relearning the weights for the deep network. In this paper, we propose a novel object detection method that can handle newer categories without the need of obtaining new labeled data and retraining the network. Our approach learns the visual similarities between various clothing categories and predicts a tree of categories. The resulting framework significantly improves the generalization capabilities of the detector to novel clothing products.