Figure 6
Details of the inferior sub-network of RetiNerveNet (best viewed in color): The sub-network has four main blocks. The first block consists of convolutional and max pooling layers, augmented by a convolutional skip connection. The second block only consists of the Recursive Progression Layer (RPL) which recursively models the outward spatial movement from the optic disc. None of the blocks except the RPL, correspond to any spatial movement along the retina. The next block consists of convolutional and max pooling layers, augmented by another convolutional skip connection. The final block consists of a single convolutional layer which combines the different representations obtained from the previous block into a scalar value. The subsequent max pooling layers bring down the dimension of the output to the intended size (5-dimensional vector). All convolutional layers are equipped with ReLU activations, except both the skip connections and the final convolutional layers of the corresponding blocks. The network respectively uses the first 5, 4, 4, 4, 4, 4, and 1 dimensions of the outputs after 7 successive recursive passes to obtain an estimate of the entire superior half of the visual field (consisting of 26 points). The visual field points estimated after successive passes are differentiated using contrasting shades.
