Input-Dependent Uncorrelated Weighting for Monte Carlo Denoising

Jonghee Back¹, Binh-Son Hua², Toshiya Hachisuka³, and Bochang Moon¹

¹ GIST, South Korea, ² Trinity College Dublin, Ireland, ³ University of Waterloo, Canada

Overview

In this interactive viewer, we have compared our kernel with input-independent kernels (a uniform kernel and a cross-weighting) and input-dependent kernel (a learning-based kernel (DC) [Back et al. 2020]) by taking two different images (path tracing with independent sampling and correlated sampling using common random numbers (CRN)) as inputs. We have compared our method with a recent biased denoiser, auxiliary feature guided self-attention (AFGSA) [Yu et al. 2021]. Since the denoiser takes only independent pixel estimates as input, we set the total sample budgets for both methods to be the same. We additionally demonstrate that our method can be used to improve the unbiased result of gradient-domain rendering [Kettunen et al. 2015] (i.e., L2 reconstruction). To quantitatively compare the quality of results, we measure the relative L2 error [Rousselle et al. 2011] with a small value 1e-2.

Comparisons with input-independent kernels

Comparisons with an input-dependent kernel

Comparisons with a recent biased denoiser

Results for gradient-domain rendering

References

[Back et al. 2020] Jonghee Back, Binh-Son Hua, Toshiya Hachisuka, and Bochang Moon. 2020. Deep Combiner for Independent and Correlated Pixel Estimates. ACM Trans. Graph. 39, 6, Article 242 (Nov. 2020), 12 pages.

[Kettunen et al. 2015] Markus Kettunen, Marco Manzi, Miika Aittala, Jaakko Lehtinen, Frédo Durand, and Matthias Zwicker. 2015. Gradient-Domain Path Tracing. ACM Trans. Graph. 34, 4, Article 123 (July 2015), 13 pages.

[Rousselle et al. 2011] Fabrice Rousselle, Claude Knaus, and Matthias Zwicker. 2011. Adaptive Sampling and Reconstruction Using Greedy Error Minimization. ACM Trans. Graph. 30, 6, Article 159 (Dec. 2011), 12 pages.

[Yu et al. 2021] Jiaqi Yu, Yongwei Nie, Chengjiang Long, Wenju Xu, Qing Zhang, and Guiqing Li. 2021. Monte Carlo Denoising via Auxiliary Feature Guided Self-Attention. ACM Trans. Graph. 40, 6, Article 273 (Dec. 2021), 13 pages.

Acknowledgements

We would like to thank nacimus (Bathroom), NovaZeeke (Classroom), cekuhnen (Coffee-maker), MrChimp2313 (House), Jay-Artist (Kitchen), UP3D (Lamp), Benedikt Bitterli (Material-testball) and Wig42 (Staircase) for providing the following test scenes. Additionally, we would like to express our appreciation to Joey Litalien, Jan Novák, and Benedikt Bitterli for the interactive viewer.