Attentional Bi-directional LSTM for Semantic Attribute Prediction

Attribute prediction is a basic task in CV field, and it belongs to a multi-label prediction problem in practical terms. Most studies using deep features to handle the problem while ignoring the potential dependencies that exist within attributes and images. Differ from previous work, we propose a novel deep architecture named ABLSTM, which not only taking advantage of CNN and Bi-LSTM, but also utilizing a multi-task-multi-loss design for attribute detection. Based on ABLSTM, we further construct a simple but extremely effective regression module to improve the accuracy of high-level abstract semantic attributes. Extensive experiments on the largest attribute prediction dataset of DeepFashion show the consistency superiority and efficiency of the proposed model.