Self-Supervised Post-Correction for Monte Carlo Denoising

Jonghee Back, Binh-Son Hua, Toshiya Hachisuka, Bochang Moon

Overview

This code is the official implementation of SIGGRAPH 2022 paper, Self-Supervised Post-Correction for Monte Carlo Denoising. For more detailed information, please refer to our project page or other materials as below:

• project page
• paper
• supplementary report
• interactive viewer

We have tested the code on Ubuntu 20.04 with NVIDIA GeForce RTX 3090 graphics card, and we have additionally checked that this code works well on the following graphics cards: NVIDIA GeForce RTX 3060, TITAN RTX, Quadro RTX 8000.

Requirements

We highly recommend running this code through Docker and Nvidia-docker on Ubuntu. Please refer to the detailed instruction for the installation of Docker and Nvidia-docker.

Usage

Building Docker image

Please use run_docker.sh to build a docker file with Dockerfile and run it.

Running self-supervised post-correction

In order to run the provided post-correction codes, you can proceed in the following order:

1. Prepare image buffers in data folder.
2. Check configuration setting in main.py.
3. Run main.py with arguments regarding a test scene.

If you have modified CUDA code in custom_op folder, please build the op library using build_custom_op.sh to update post_correction_lib.so file.

Example code

Please run this command for the provided example data as follows:

python main.py --scene dragon --spp 128 --deno afgsa

Example data

The example data consists of input-denoised pairs with auxiliary features rendered by PBRT-V3. We provide four test scenes (bathroom, hair, dragon and sanmiguel) with three different Monte Carlo denoisers (Kernel-Predicting Convolutional Networks for Denoising Monte Carlo Renderings (KPCN), Adversarial Monte Carlo Denoising with Conditioned Auxiliary Feature Modulation (AMCD) and Monte Carlo Denoising via Auxiliary Feature Guided Self-Attention (AFGSA)). The example data can be available here:

• Example data (bathroom)
• Example data (hair)
• Example data (dragon)
• Example data (sanmiguel)

After downloading and unziping the data, please place it in data folder.

License

All source codes are released under a BSD License.

Citation

@inproceedings{Back22,
author = {Back, Jonghee and Hua, Binh-Son and Hachisuka, Toshiya and Moon, Bochang},
title = {Self-Supervised Post-Correction for Monte Carlo Denoising},
year = {2022},
isbn = {9781450393379},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3528233.3530730},
doi = {10.1145/3528233.3530730},
booktitle = {ACM SIGGRAPH 2022 Conference Proceedings},
articleno = {18},
numpages = {8},
keywords = {self-supervised denoising, self-supervised learning, self-supervised loss, Monte Carlo denoising},
location = {Vancouver, BC, Canada},
series = {SIGGRAPH '22}
}

Contact

If there are any questions, issues or comments, please feel free to send an e-mail to jongheeback@gm.gist.ac.kr.

Credits

We have used EXR I/O functionalities (exr.py in codes/image_io) from Kernel-Predicting Convolutional Networks for Denoising Monte Carlo Renderings (KPCN) project.