Deferred rendering methodologies analysis and comparison

Real-time photorealistic rendering has always been one of the long-standing goals in computer graphics. In particular, this objective often goes down to low-latency photorealistic two-dimensional image generation based on analytic three-dimensional environment description and spectator location and view parameters using a typical personal computer with consumer graphics hardware with relatively limited computing capability. The fundamental problems one has to solve to perform rendering are main view visibility determination and shading, i. e. finding the environment surface element that is observed in the particular pixel of the resulting image and calculating the amount of light this surface element reflects in the observer's direction for all image's pixels. Due to the computation capability constraints of the target hardware, it is crucial to perform these operations as efficiently as possible. This work is aimed at analysis and experimental comparison of the modern methodologies used to tackle main view visibility determination and shading tasks. Due to modern lighting calculation methods' complexity and high geometric detail of the virtual environments, forward rendering has become impractical as a general-purpose rendering approach and only remains in use when specific geometry or material types are involved. Deferred rendering with G-buffer generation has become a state-of-the-art solution for products demanding high visual quality and fidelity. G-buffer generation can be performed in different ways including rasterized G-buffer generation and visibility buffer method. A theoretical overview and comparison of these techniques are presented in this work. Also, we use a demonstration application we implemented to perform an experimental comparison of G-buffer generation techniques in various conditions. Our experimental results can be used as a guidance when designing production rendering solutions.