*5.3. Discussion*

We summarize our two major finding based on our experimental results. First, our framework can successfully generate 3D images of porous pavement microstructure from a single 2D image using 3D-IWGAN with enhanced GP. Figures 10 and 11 illustrate the generated 3D microstructure images as a visualization. Second, the generated 3D images are realistic in terms of the physical properties of permeable pavement extracted from our generated images. From Figure 12 and Table 5, we observe that the computed values of porosity and surface area fall in the correct ranges of the standard values. Figure 13 also displays a low error value of about 3.54% on average in nine different samples.

There are some limitations that restrain us from obtaining 100% accuracy despite getting decent results in 3D image generation and the computations of physical properties. Some limitations arise from the limited 2D image dataset as the input whereas others arise from the computational resources for our proposed framework. To provided a 2D input image for our framework, scanned images of pavement microstructure are needed at the pre-processing step. Since we only had a limited set of scanned images, we applied image resampling to obtain enough 2D images for our experiments. Powerful computational resources are important for generating 3D images with high resolution. For future experiments, several machines can be utilized as a distributed way to obtain large 3D images with higher resolutions.
