Figure 1.

Structure of the 3D classifier network combining a single convolutional layer (kernel $3\times 3\times 3$, 8 channels) with a down-convolutional layer (kernel $3\times 3\times 3$, 8 channels, stridding $2\times 2\times 2$) as the main building block. The overall network stacks 4 of these main building blocks followed by two fully connected layers (16 and 2 units) with totally 0.3 million trainable parameters. Each layer is followed by a Rectified Linear Unit (ReLU) nonlinearity, except for the output layer where a Softmax activation is applied. Dimensionalities between layers describe the tensor size after each network layer.