Encoding Spatial Distribution of Convolutional Features for Texture Representation