DeepAO: Efficient Screen Space Ambient Occlusion Generation via Deep Network

Abstract

Ambient occlusion (abbr. AO) plays an important role in realistic rendering applications because AO produces more realistic ambient lighting, which is achieved by calculating the brightness of certain screen parts based on objects' geometry. However, the baseline computation of AO algorithm is time-consuming, which limits its application for real-time rendering. Currently, most AO algorithms are based on screen space to reduce the computational consumption, which leads to unrealistic results due to the usage of artificial features. To overcome these challenges, in this paper, we first create a well-crafted dataset with the pair of deferred shading buffer data and ground-truth AO shaded images. Then, we design an efficient deep neural network for the screen space AO image generation, based on which we further design a Compute Shader Library to compute the shaded AO images. Our extensive experimental results show that our method achieves competent performance than existing screen space ambient or volumetric ambient based AO methods both in visual quality and efficiency. INDEX TERMS Ambient occlusion, rendering, shading, deep neural network.

Pyramid HBAO — a Scalable Horizon-based Ambient Occlusion Method

“…Therefore, later appeared the works [15,16], which use a deeper U-Net-like architecture of the Encoder-Decoder type, which allows the network to extract more complex features by creating feature maps of different scales.…”

AOGAN: A generative adversarial network for screen space ambient occlusion

“…The dataset was provided by Nalbach et al [7] and Zhang et al [21], and consists of 105,000 pairs of deferred shading G-buffer data for 27 scenes, with corresponding reference images. We used 90,000 pairs of images to train the network, 10,000 for verification, and the remaining 5000 for testing.…”

“…Since the generator itself cannot converge quickly, as shown in Fig. 7 We applied the rendering framework proposed by Zhang et al [21] to further improve the network rendering time from 12.5 to 3.5 ms. Comparative results are provided in Table 3.…”

Accounting for Non-Uniform Ambient Lighting in Ray Traced Ambient Occlusion

“…Another promising direction could be to use neural networks to predict the ambient occlusion. Zhang et al [12] recently proposed one such technique with a detailed comparison to existing methods.…”