Final Project: Choose your Topic

This assignment is to choose your topic for your final project. You will need to submit your choice by the deadline for this assignment. Doing so is worth 5% of the score of your final project.

Note: you may work in groups of two. If you do so, indicate your project partner in your submission to this assignment. Both teammates need to send in a submission for this assignment with their selection of topic and partner name.

There are two main options for project topics:

1. Use your own research

You may choose to use work on your own research as the final project. This is possible if your research is be related to rendering, 3D geometry, or deep learning for visual data (images, video). If you choose this option, submit a short description (half a page to a page) of your research, where you clearly indicate

• the problem statement,
• the current status of the research,
• and the work that you will do for the final project in this class. This is what you will be graded on.

2. Select a topic from a proposed list

Alternatively, you may select a topic from the list below. In this case, simply submit a text file indicating your choice. Include any questions you may have about the topic.

Note: for topics related to deep learning you need access to a CUDA-capable GPU.

1. Extend Nori by implementing scattering in participating media. You could start with single scattering as discussed in class. This topic is also described in the PBRT book Links to an external site., Chapter 15-15.3. Evaluate your implementation by building a nice scene to demonstrate the new functionality.
2. Extend Nori by implementing subsurface scattering. You can build on the high-level discussion in class, the material in the PBRT book Links to an external site., Chapter 15.4 and 15.5, and the original research article Links to an external site.. Evaluate your implementation by building a nice scene to demonstrate the new functionality.
3. Extend Nori by implementing a bidirectional light transport algorithm, such as bidirectional path tracing, photon mapping, or Metropolis light transport. These techniques are described in the PBRT book Links to an external site. in Chapter 16.  Bidirectional path tracing is also described in the document Download document
   that was provided earlier in the class, and in Eric Veach's PhD thesis Links to an external site., Chapter 10. Evaluate your implementation by comparing the rendering performanc (in terms of noise) to the unidirectional path tracer. If you choose this topic, indicate which technique (bidirectional path tracing, photon mapping, or Metropolis light transport) you want to implement.
4. Implement a deep learning based denoiser for images rendered with Nori by building on the "Deep Image Prior Links to an external site."by Ulyanov et al. The idea of this work is to optimize a convolutional neural network on a single image, and they show that this can be used as a prior for various image processing tasks including denoising. This is interesting, because you do not need an image database to train on. The goal of this topic is to extend the deep image prior to Monte Carlo denoising, by providing additional input channels (per pixel normals, texture, depth), similar as is done in other Monte Carlo denoising approaches. To evaluate the results, one could compare the original deep image prior approach (without additional inputs) to the extended version (with normals, texture, depth as additional inputs) to show the improvement due to the additional inputs.
5. Implement a deep learning based denoiser for images rendered with Nori by building on the "Noise2Noise: Learning Image Restoration without Clean Data Links to an external site." by Lehtinen et al. This work shows that it is possible to train a denoiser without clean, noise-free ground truth data. This is interesting for Monte Carlo rendering, because it means you do not need to render high-quality, noise-free reference images to train on. The goal of this topic is to implement a denoising network for Nori and train it without ground truth data to show the feasibility of this approach. To evaluate the approach, create a training and testing dataset of noisy images from different scenes and collect quality measurements. 
6. This is more adventurous: Implement and improve a neural importance sampling technique in Nori. You would follow this Links to an external site. and this Links to an external site. paper. This might be more difficult than the other topics, and we would provide additional guidance if you choose this.