基於幾何感知重建法之眼神局部光照近似系統

Abstract

近幾年，在虛擬實境（Virtual Reality）以及擴增實境（Augmented Reality）技術日漸火紅之下，高解析度以及高刷新率的渲染系統需求使得視點渲染技術（Foveated Rendering）也日漸吸引研究者們的注意。此方法係一種利用人眼視覺系統的特性，透過將運算資源集中在人眼視點中心（Fovea）以保持視線中心清晰度，同時去降低視線外圍區域（Periphery）的影像質量，進而減少整體渲染花費的運算方法。 許多方法專注於將幾何圖形做次取樣後以較低的解析度進行片元渲染。然而，由於接續於其後的螢幕空間升取樣往往在視線外圍區域帶來使用者能觀察到的模糊效果，使得這些方法無法更進一步地降低著色率。再者，目前大部分的先進實時渲染系統都具備演算全局光照（Global Illumination）的能力，這在視點渲染技術裡並沒有被周全考慮過。 本篇論文引出了視點渲染以及全局光照之間的相似處，並且提出了第一套使用幾何感知重建法來近似光照的視點渲染系統。我們主要著重於常常佔據渲染效能的全局光照。其核心為一個由緩衝器組成的金字塔結構，並儲存根據使用者視線位置取樣的幾何資訊。透過這個金字塔結構，我們能將漫射全局光照從一個低解析度的著色圖重建回來，並且保有視線中心的清晰度。最終，直接光照與間接光照將會被融合成最終的畫幀。實驗結果顯示，我們所提出之演算法能達到6倍至8倍的加速，同時保有最小的視覺損失。

Recently, while applications of Virtual Reality and Augmented Reality (VR/AR) are getting popular, foveated rendering, a rendering schema for reducing rendering cost by utilizing properties of the human visual system, is also getting more attention especially for high-resolution and high-refresh-rate VR/AR systems. It reduced rendering time by allocating primary of computational resources to the fovea for fidelity preservation and decreasing image quality of the periphery. Many methods have focused on subsampling geometry and run fragment shading at a relatively low rate. However, they cannot reduce the shading rate more aggressively since those following screen-space upsampling processes would bring a blurred result at the peripheral region. This artifact would usually make rendering result distinguishable from the original for users. Moreover, a large portion of state-of-the-art real-time rendering systems involves real-time global illumination, which is not fully considered in existing foveated rendering algorithms. In this thesis, we draw an analogy between foveated rendering and global illumination and propose the first foveated rendering algorithm base on a geometry-aware reconstructing strategy to approximate illumination. We put our emphasis on the computation of global illumination which usually dominates rendering performance. Our key component is a pyramid of buffers storing geometry information which is sampled according to the gaze position of users. With this pyramid, we can reconstruct diffuse global illumination from a low-resolution shaded image while keeping fidelity at the fovea. In the end, direct illumination is blended with indirect illumination to form an entire frame. Our experiments show that we can get 6x-8x speedup in rendering time with a minimal perceptual loss-of-detail.