NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
Ben Mildenhall*¹, Pratul P. Srinivasan*¹, Matthew Tancik*¹, Jonathan T. Barron², Ravi Ramamoorthi³, Ren Ng^1
1UC Berkeley, ²Google Research, ³UC San Diego
*denotes equal contribution

A PyTorch re-implementation of Neural Radiance Fields.

Since the original release, and the concurrent pytorch implementation wrote all the codes basically in just two files, run_nerf.py and run_nerf_helper.py, which makes it very hard to read, along with many unnecessary operations, I re-organize the whole code architecture as below, which is more concise and readable.

├── core
│   ├── datasets
│   │   ├── create_dataloader.py
│   │   ├── create_dataset.py
│   │   ├── create_rays.py
│   │   ├── load_blender.py
│   │   ├── load_deepvoxels.py
│   │   └── load_llff.py
│   ├── models
│   │   ├── create_model.py
│   │   ├── embedder.py
│   │   └── nerf.py
│   └── utils
│       ├── create_configs.py
│       ├── metrics.py
│       ├── test_nerf_utils.py
│       └── train_nerf_utils.py
├── run_nerf.py

Besides the re-organization, I re-implement some of the core functions in NeRF:

• core/datasets/dataset.py: a unified dataset for blender, llff data type, with all N*H*W rays generated;
• core/datasets/dataloader.py: a unified dataloader for both no_batching=True/False;
• core/datasets/create_rays.py: a vectorized implementation of rays generation, N*H*W rays once;
• core/models/embedder.py: a vectorized implementation of positional encoding;

With all the modifications describes above, this NeRF version is a little🫠 faster than the concurrent pytorch implementation, which may illustrates the bottleneck may relies on the caching of data, according to the re-implementaion by krrish94.

Download data for two example datasets: lego and fern

bash download_example_data.sh

To train a low-res lego NeRF:

python run_nerf.py --config configs/lego.txt

To train a high-res lego NeRF:

python run_nerf.py --config configs/lego-official.txt

To train a low-res fern NeRF:

python run_nerf.py --config configs/fern.txt

@misc{mildenhall2020nerf,
    title={NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis},
    author={Ben Mildenhall and Pratul P. Srinivasan and Matthew Tancik and Jonathan T. Barron and Ravi Ramamoorthi and Ren Ng},
    year={2020},
    eprint={2003.08934},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}