FSA-Net: Learning Fine-Grained Structure Aggregation for Head Pose Estimation From a Single Image | Synapse