Learning to Refine Object Contours with a Top-Down Fully Convolutional Encoder-Decoder Network.