Runner structure overhaul

In this release, we bring significant changes to runners for training and rendering neural radiance fields.

Previously, the runners (runners/run_train.py, runners/run_render.py) were not written with scalability in mind.
This hindered the extension of codebases as training/rendering scripts were dependent on model structure, training strategy, etc.

To improve the scalability of torch-NeRF, we built a new structure for runners from scratch, enabling model-agnostic initialization and code execution.

For details, please refer to #15 and its commit history.

Full Changelog: v1.1.0...v1.2.0

Add support for LLFF dataset

This release includes:

• Abstract dataset class and helper functions for loading LLFF dataset

Most of the codes were adopted from the official TensorFlow implementation of NeRF (ECCV 2020), but we added documentation to help users understand the functionalities of codes and intentions behind them.

Disclaimer) The current release also has several (known) issues:

1. Training with NDC projection is not working. However, you can still produce nice-looking images on Blender and LLFF datasets;
2. The helper function spherify_poses in load_llff.py is still undeciphered. Meanwhile, it is not invoked while loading LLFF dataset so the current execution flow has nothing to do with it. We will add the documentation ASAP;

Currently, we are considering interesting future directions that will expand what this code base is capable of, please stay tuned for updates!

Full Changelog: v1.0.0...v1.1.0

Base NeRF model and volume rendering pipeline

This release includes:

1. A fully-functioning volume renderer implemented using Pytorch consists of camera, ray sampler, and integrator modules capable of rendering (neural) radiance fields
2. An implementation of MLP model proposed in NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis, Mildenhall et al., (ECCV 2020, Best paper honorable mention)
3. Scripts for training and visualizing the neural radiance fields
4. User friendly configuration file interface based on Hydra

This version of codebase implements the key ideas, with detailed documentations, proposed in the paper such as:

1. stratified sampling
2. density-based hierarchical sampling
3. positional encoding

Disclaimer) However, this release does NOT include or reproduce:

1. dataset and loaders for LLFF dataset
2. scripts for quantitative evaluations
3. quantitative metrics claimed in the paper

Please stay tuned for future releases that are highly likely to include the components mentioned above and more fancy stuffs!

Full Changelog: https://github.com/DveloperY0115/torch-NeRF/commits/v1.0.0