Guided Visual Attention Model Based on Interactions Between Top-down and Bottom-up Information for Robot Pose Prediction.