Computing and Visualizing Image-level One-Order Facial Attribute Relations with Deep Features

摘要

Researches have shown that Multi-label classification performance could be improved by using spatial dependency [1-3] and label relations [1,4-6]. In the face attributes prediction task, spatial dependency is minor to label relations because some attributes usually do not have spatial dependency (attractive, Male, etc.). In utilizing label relation, however, little research extract relations for each image. In this paper, we not only extract the image-wise label relation, but also decompose the relation map to the general relations and image specified relations. The extracted relation map is one-order and could be visualized easily to give better understanding on face attribute prediction. Experiments also show that the proposed sub-network increases prediction performance.