Fig. 2.

Structure of the networks. Due to the pooling layers the network works on four different resolutions. This allows a large receptive field at low memory cost during training. All convolutional layers use rectified linear unit activations apart from the last one using a softmax activation to produce the final output probabilities