HIGHLY PRODUCTIVE NEURAL SHADING OF THREE-DIMENSIONAL FIGURES USING PIX2PIX MODEL

Abstract

In the article the generative neural networks-based two-stage system for three-dimensional figures shading is developed. The advantages and disadvantages of standard three-dimensional images rendering approaches are analyzed. The features of neural rendering are analyzed. The neural research directions of generating shaded images from figures’ geometrical data and two-dimensional sketches, getting geometrical data from images are examined. The features of architecture and usage of generative adversarial networks are described. The need in the development of new neural rendering methods for increasing the productivity of three-dimensional figures’ surfaces shading is justified. The proposed system for figures neural shading that contains Pix2Pix models for images formation and improving their quality is described. The development of dataset based on ShapeNet figures array for neural network training is described. The proposed volumetric representation of figures’ vertices information that is used as neural system input is examined. The architectures of generator and discriminator of Pix2Pix for figures shading are described. The information about neural network training duration and used error functions is provided. The plots of discriminator’s and generator’s error functions changes during the Pix2Pix training for figures shading are built. Using the SSIM metric and test figures dataset the image generation quality level is evaluated. The architectures of generator and discriminator of Pix2Pix for improving the quality of generated images and their scaling are described. The plots of generator’s and discriminator’s error metrics changes during the training of Pix2Pix for images quality improvement are built. The examples of generated by two-stage neural system shaded figures images are provided. Using the SSIM metric the quality of generated by the second system’s stage images is evaluated. The speed of figures shading using the proposed system and render Blender Eevee are compared. The developed neural system allows to generate realistic images and improve the productivity of figures’ surfaces shading.