Efficient Convolutional Patch Networks for Scene Understanding