Attribute based approach for clothing recognition

Abstract

Clothing recognition is hot topic for its potential benefits to lots of visual tasks, such as people identification, pose estimation and recommendation system. However, due to the wide variations of clothing appearance and the “semantic gap” between low-level features and high-level category concepts, clothing recognition is very challenging. To narrow this gap, a novel method, which uses intermediate attributes to bridge low-level features and high-level category labels, is proposed. This method first recognizes local attributes from low-level visual features, and then infers clothing category based on these attributes. To this end, DPM models and pixel-level parsing are applied to obtain geometric structure attributes, such as collar shape, and geometric size attributes, such as sleeve length, respectively. Then, Multiple Output Neural Networks are built to predict clothing category based on attributes. Experiments show that the performance of our method is superior to two stateof- the-art approaches on both of attribute and category recognition.