A Pipeline for NeRF Experimentation and Visualization

The goal for this project is to develop a NeRF model training pipeline within the OpenCV library. The end result will be a fully functioning data processing pipeline that is well integrated with the NeRF model training procedure, as well as online documentation that detail how to use the pipelines properly. There are 3 key deliverables for this project. First, implement data loaders that can process user-taken images of real-world scenes. This will involve using the COLMAP and OpenCV data processing pipelines in order to extract the camera poses and epipolar geometry of the scene; these are requirements for the model's ability to learn a meaningful representation of the given scene. Second, build, train, and test the actual NeRF model. This will involve using PyTorch's built-in CUDA support and extensive usage of the PyTorch library as a whole in order to incorporate critical components of the NeRF architecture such as positional encoding, ray sampling, and regularization into our model design. This step will also include rendering procedures that will allow users to extract images/videos, depth maps, and other logging information such as the training loss curve from the fully trained model. Third, create simple and intuitive documentation on how to use both the data loaders and the NeRF code base. This will involve creating a page on OpenCV's documentation (via Doxygen) that will list out specific terminal commands that the user should use to train the model. It will also include a summary of required packages, recommended environment configurations, and links to other helpful resources such as YouTube video demonstrations, which will all be geared towards increasing the usability and accessibility of the pipeline.