Attributes-oriented clothing description and retrieval with multi-task convolutional neural network

Abstract

This paper seek answer to question how to search clothing when consumer pays attention to a part of clothing. A novel framework is proposed to solve above problem by attributes. First of all, Fast-RCNN detects person from complex background. Then a Convolutional Neural Network (CNN) is combined with Multi-Task Learning (MTL) to extract features related to attributes. Next Principal Component Analysis (PCA) reduce dimensionality of feature from CNN. Finally, Locality Sensitive Hashing (LSH) searches similar samples in the gallery. Extensive experiments were done on the clothing attribute dataset, experimental results proves this framework is effective.