Figure 3.

Overview of multi-scale VGG Network Architecture. Image patches of size 224 and 112 are provided as input to two separate network streams, from which feature vectors of 512 are concatenated and used to make a classification decision over the four classes. Numbers on top of features maps indicate image size at the corresponding cross section, and the legend in the top right displays the number of filter channels in each color coded layer. Batch normalization is added between every convolutional layer, as well as after the feature concatenation layer, and ReLU activation is used between all layers until the final classification.