CFAM: Estimating 3D Hand Poses from a Single RGB Image with Attention