Figure 2:

Illustration of the 3D deconvolutional layer and voxel deconvolutional layer. For convenience, we use number 0 to denote the input, and number 1 – 8 to denote the eight intermediate feature maps which are generated from convolutional operations. The input size is 2 × 2 × 2 and the desired out size is 4 × 4 × 4. The size of each intermediate feature map is also 2×2×2. (a) shows how the intermediate feature maps are generated independently from eight different kernels in a regular 3D deconvolutional layer. This independent generation process is also indicated in the lower orange bar in (c). (b) shows how the intermediate feature maps are generated sequentially in a voxel deconvolutional layer. The i^th feature map are built on the 1^st, …,(i −1)^th feature maps. In iVoxelDCLa and iVoxelDCLc layers, input data is also included in this process. This sequential generation process is also indicated in the lower blue bar in (c). (c) displays the periodically shuffling and combination process of eight feature maps to form the final output of a deconvolutional layer.