AO-Net: Efficient Neural Network for Ambient Occlusion

jiayi wang, Fan Zhou, xiang zhou, yongle li, Xueqian Wang

Keywords: Ambient occlusion, Neural networks, Global illumination, Shading

TL;DR: Efficient Neural Network for Ambient Occlusion

Abstract: Screen space based ambient occlusion is widely applied in real-time 3D applications due to its high efficiency, however, it frequently exhibits artifacts including banding and blurring. In this paper, we propose AO-Net, a learning-based method for fast and high-quality ambient occlusion generation. Our neural network is built upon kernel prediction-based architecture with careful input screen space feature selection, leading to a light-weight and compact solution. Experiment results indicate that our approach can achieve visual quality to a level in comparable with ray-traced solutions, meanwhile maintaining real-time performance. In addition, our method can be easily integrated into existing rendering pipelines and shows robustness for unseen scenes.

Track: Graphics

Summary Of Changes:

We thank the reviewers for the invaluable comments that helped us improve our submission “AO-Net: Efficient Neural Network for Ambient Occlusion”. We revised the paper according to the issues and suggestions raised by the reviewers. We hope that the new material will lift all concerns.

[Meta review] "rerun the study while reserving some scenes strictly for testing"
We strictly reserved scenes for testing during the experiment. The training and test scenes we use are disjoint and this is the case both in our dataset and in the deepao dataset. We have supplemented this in the paper, see Section 4.1 (first paragraph).

[Meta review] “add technical discussions on the results of Table 3 that suggest including the depth as an additional input to the U-Net is useless”
Normal is the cross product of the gradient of the depth buffer in the smooth region, so normal can be seen as knowledge distilled from depth, and is sufficient to express the geometric information required for AO generation. More importantly, because the depth of different scenes varies greatly, the quality can degrade greatly if the depth is not normalized or is not normalized correctly. In practice, adding depth will lead to difficult network convergence, thus we consider the instability of depth information to play a negative role in AO generation tasks. We added technical discussions on the results of Table 3 in Section 4.2.2 (second paragraph).

[Reviewer] “the illustration of dataset generation used in the examination is too simple. ”
We have added more detailed illustrations of dataset generation in Section 4.1 (first paragraph), including the number and categories of scenes in the dataset.

[Reviewer] “Are there scenes where the method does not work as well? If so, that could point to future work.”
However, our work has the same limitations as all screen space solutions. G-buffer does not contain complete information about the scene, which may lead to defects on screen boundaries or special perspectives. We clarified this in the second paragraph of Section 5.

[Reviewer] “In Table 1, the runtime of GTAO is given as "-". What does this mean?”
"-" is not that the method requires zero time, but that we did not measure the run time of the GTAO. Partly because we use GTAO in Blender for comparison, it is difficult to measure the running time of the process in Blender. On the other hand, current learning-based methods have no obvious advantage over traditional screen space AO methods in terms of speed, so we mainly focus on speed compared with other learning-based methods. An explanation of this has been added in Section 4.2 (first paragraph).

[Reviewer] “The paper would benefit from an editing pass for typos and grammar. Try to fix up the bibliography. ”
We have corrected the typos and grammar and revised the references.

We hope that our changes answer to all concerns. We thank the reviewers again for their careful reviews of our submission and believe that this input has helped us to substantially improve the submission. Hopefully our work will provide a significant benefit to the community and spark new ideas.