This repository contains the code and data related to the thesis "Denoising Monte-Carlo Noise with Deep Learning: A Rodent-Specific Approach". The thesis investigates neural network architectures and learning methods used for denoising images generated by Monte-Carlo renderers, specifically using the high-performance renderer Rodent. The primary objective is to evaluate the performance of deep learning techniques in reducing Monte-Carlo noise, with a focus on creating noise-free images efficiently and without large-scale computational resources.
Monte-Carlo rendering, particularly through path tracing, produces high-quality images but suffers from significant noise, especially in scenarios involving global illumination. While various techniques exist to reduce this noise, deep learning has emerged as a powerful tool to achieve denoising with minimal computational effort.
This thesis introduces a Rodent-specific denoising approach where neural networks are trained to remove Monte-Carlo noise from images rendered by Rodent. This repository includes the following components:
- Rodent Renderer: The high-performance renderer used to generate noisy images for training and testing.
- Denoising Framework: A deep learning-based framework developed to denoise images, leveraging architectures and strategies tailored for Monte-Carlo noise.
For a comprehensive explanation of the methods, theoretical background, and experimental results, please refer to the thesis document included in this repository: Thesis: ./thesis/JunkawitschThesis.pdf
- Rodent Renderer: ./rodent/
- Denoising Framework: ./mcdenoise/
The training and validation data is not included in this repository. You can contact me if you want to have access.
First and foremost, I want to thank Prof. Dr. Ing. Philipp Slusallek and the computer graphics chair at Saarland University for the great opportunity to work on this topic. Furthermore, I want to thank my advisor Ömercan Yazici, M.Sc., for his patient support and valuable insight into different important key points. Additionally, I want to acknowledge Joshua Meyer, who concurrently worked at the efficient implementation of the denoising approach in Rodent and was always willing to discuss various ideas.