Cast Shadow Generation Using Generative Adversarial Networks

We propose a computer graphics pipeline for 3D rendered cast shadow generation using generative adversarial networks (GANs). This work is inspired by the existing regression models as well as other convolutional neural networks such as the U-Net architectures which can be geared to produce believable global illumination effects. Here, we use a semi-supervised GANs model comprising of a PatchGAN and a conditional GAN which is then complemented by a U-Net structure. We have adopted this structure because of its training ability and the quality of the results that come forth. Unlike other forms of GANs, the chosen implementation utilises colour labels to generate believable visual coherence. We carried forth a series of experiments, through laboratory generated image sets, to explore the extent at which colour can create the correct shadows for a variety of 3D shadowed and un-shadowed images. Once an optimised model is achieved, we then apply high resolution image mappings to enhance the quality of the final render. As a result, we have established that the chosen GANs model can produce believable outputs with the correct cast shadows with plausible scores on PSNR and SSIM similarity index metrices.