That means we add subsequent convolutional layers to our VGG-16 architecture thus enabling to extract features at multiple scales and progressively decrease the size of the input to each subsequent layer.