Dai et al. use a 3D convolutional neural network architecture called 3D-Encoder-Predictor Network for shape completion. Figure 1 illustrates the high-level pipeline. The first step consists of two networks which are combined in the framework of their 3D Encoder-Preodictor Network as illustrated in Figure 2. The input is a two channel volume encoding the signed truncated distance function (STDF) and the output is only a distance function (DF). Nearest neighbors of the output shape (in resolution $32^3$) are searched utilizing features taken from a 3D classification network following . Finally, the output volume and the nearest neighbor shapes are used to produce a higher-resolution mesh ($128^3$), see the paper for details.
Figure 1 (click to enlarge): High-level illustration of the proposed approach as described in the text.