iNeRF: Inverting Neural Radiance Fields for Pose Estimation
Lin Yen-Chen*1,2,
Pete Florence*1,
Jonathan T. Barron1,
Alberto Rodriguez2,
Phillip Isola2,
Tsung-Yi Lin1
1Google, 2MIT3
A PyTorch re-implementation of Neural Radiance Fields.
The current implementation is blazing fast! (~5-9x faster than the original release, ~2-4x faster than this concurrent pytorch implementation)
What's the secret sauce behind this speedup?
Multiple aspects. Besides obvious enhancements such as data caching, effective memory management, etc. I drilled down through the entire NeRF codebase, and reduced data transfer b/w CPU and GPU, vectorized code where possible, and used efficient variants of pytorch ops (wrote some where unavailable). But for these changes, everything else is a faithful reproduction of the NeRF technique we all admire :)
The NeRF code release has an accompanying Colab notebook, that showcases training a feature-limited version of NeRF on a "tiny" scene. It's equivalent PyTorch notebook can be found at the following URL:
https://colab.research.google.com/drive/1rO8xo0TemN67d4mTpakrKrLp03b9bgCX
A neural radiance field is a simple fully connected network (weights are ~5MB) trained to reproduce input views of a single scene using a rendering loss. The network directly maps from spatial location and viewing direction (5D input) to color and opacity (4D output), acting as the "volume" so we can use volume rendering to differentiably render new views.
Optimizing a NeRF takes between a few hours and a day or two (depending on resolution) and only requires a single GPU. Rendering an image from an optimized NeRF takes somewhere between less than a second and ~30 seconds, again depending on resolution.
To train a "full" NeRF model (i.e., using 3D coordinates as well as ray directions, and the hierarchical sampling procedure), first setup dependencies.
In a new conda
or virtualenv
environment, run
pip install -r requirements.txt
Use the provided environment.yml
file to install the dependencies into an environment named nerf
(edit the environment.yml
if you wish to change the name of the conda
environment).
conda env create
conda activate nerf
Once everything is setup, to run experiments, first edit config/lego.yml
to specify your own parameters.
The training script can be invoked by running
python train_nerf.py --config config/lego.yml
Optionally, if resuming training from a previous checkpoint, run
python train_nerf.py --config config/lego.yml --load-checkpoint path/to/checkpoint.ckpt
An optional, yet simple preprocessing step of caching rays from the dataset results in substantial compute time savings (reduced carbon footprint, yay!), especially when running multiple experiments. It's super-simple: run
python cache_dataset.py --datapath cache/nerf_synthetic/lego/ --halfres False --savedir cache/legocache/legofull --num-random-rays 8192 --num-variations 50
This samples 8192
rays per image from the lego
dataset. Each image is 800 x 800
(since halfres
is set to False
), and 500
such random samples (8192
rays each) are drawn per image. The script takes about 10 minutes to run, but the good thing is, this needs to be run only once per dataset.
NOTE: Do NOT forget to update the
cachedir
option (underdataset
) in your config (.yml) file!
A Colab notebook for the full NeRF model (albeit on low-resolution data) can be accessed here.
Once you've trained your NeRF, it's time to use that to render the scene. Use the eval_nerf.py
script to do that. For the lego-lowres
example, this would be
python eval_nerf.py --config pretrained/lego-lowres/config.yml --checkpoint pretrained/lego-lowres/checkpoint199999.ckpt --savedir cache/rendered/lego-lowres
You can create a gif
out of the saved images, for instance, by using Imagemagick.
convert cache/rendered/lego-lowres/*.png cache/rendered/lego-lowres.gif
This should give you a gif like this.
All said, this is not an official code release, and is instead a reproduction from the original code (released by the authors here).
The code is thoroughly tested (to the best of my abilities) to match the original implementation (and be much faster)! In particular, I have ensured that
- Every individual module exactly (numerically) matches that of the TensorFlow implementation. This Colab notebook has all the tests, matching op for op (but is very scratchy to look at)!
- Training works as expected (for Lego and LLFF scenes).
The organization of code WILL change around a lot, because I'm actively experimenting with this.
Pretrained models: Pretrained models for the following scenes are available in the pretrained
directory (all of them are currently lowres). I will continue adding models herein.
# Synthetic (Blender) scenes
chair
drums
hotdog
lego
materials
ship
# Real (LLFF) scenes
fern
Feel free to raise GitHub issues if you find anything concerning. Pull requests adding additional features are welcome too.
nerf-pytorch
is available under the MIT License. For more details see: LICENSE and ACKNOWLEDGEMENTS.
A shoutout to Krishna Murty for his awesome scalable PyTorch implementation, which this code base is pretty much entirely based on.