BARF 🤮: Bundle-Adjusting Neural Radiance Fields

Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey
IEEE International Conference on Computer Vision (ICCV), 2021 (oral presentation)

Project page: https://chenhsuanlin.bitbucket.io/bundle-adjusting-NeRF
Paper: https://chenhsuanlin.bitbucket.io/bundle-adjusting-NeRF/paper.pdf
arXiv preprint: https://arxiv.org/abs/2104.06405

We provide PyTorch code for all experiments: planar image alignment, NeRF/BARF on both synthetic (Blender) and real-world (LLFF) datasets, and a template for BARFing on your custom sequence.

---

Prerequisites

• Note: for Azure ML support for this repository, please consider checking out this branch by Stan Szymanowicz.

This code is developed with Python3 (python3). PyTorch 1.9+ is required.
It is recommended use Anaconda to set up the environment. Install the dependencies and activate the environment barf-env with

conda env create --file requirements.yaml python=3
conda activate barf-env

Initialize the external submodule dependencies with

git submodule update --init --recursive

---

Dataset

• Synthetic data (Blender) and real-world data (LLFF)

  Both the Blender synthetic data and LLFF real-world data can be found in the NeRF Google Drive. For convenience, you can download them with the following script: (under this repo)

  # Blender
  gdown --id 18JxhpWD-4ZmuFKLzKlAw-w5PpzZxXOcG # download nerf_synthetic.zip
  unzip nerf_synthetic.zip
  rm -f nerf_synthetic.zip
  mv nerf_synthetic data/blender
  # LLFF
  gdown --id 16VnMcF1KJYxN9QId6TClMsZRahHNMW5g # download nerf_llff_data.zip
  unzip nerf_llff_data.zip
  rm -f nerf_llff_data.zip
  mv nerf_llff_data data/llff

  The data directory should contain the subdirectories blender and llff. If you already have the datasets downloaded, you can alternatively soft-link them within the data directory.

• Test your own sequence!

  If you want to try BARF on your own sequence, we provide a template data file in data/iphone.py, which is an example to read from a sequence captured by an iPhone 12. You should modify get_image() to read each image sample and set the raw image sizes (self.raw_H, self.raw_W) and focal length (self.focal) according to your camera specs.
  You may ignore the camera poses as they are assumed unknown in this case, which we simply set to zero vectors.

---

Running the code

• BARF models

  To train and evaluate BARF:

  # <GROUP> and <NAME> can be set to your likes, while <SCENE> is specific to datasets
  
  # Blender (<SCENE>={chair,drums,ficus,hotdog,lego,materials,mic,ship})
  python3 train.py --group=<GROUP> --model=barf --yaml=barf_blender --name=<NAME> --data.scene=<SCENE> --barf_c2f=[0.1,0.5]
  python3 evaluate.py --group=<GROUP> --model=barf --yaml=barf_blender --name=<NAME> --data.scene=<SCENE> --data.val_sub= --resume
  
  # LLFF (<SCENE>={fern,flower,fortress,horns,leaves,orchids,room,trex})
  python3 train.py --group=<GROUP> --model=barf --yaml=barf_llff --name=<NAME> --data.scene=<SCENE> --barf_c2f=[0.1,0.5]
  python3 evaluate.py --group=<GROUP> --model=barf --yaml=barf_llff --name=<NAME> --data.scene=<SCENE> --resume

  All the results will be stored in the directory output/<GROUP>/<NAME>. You may want to organize your experiments by grouping different runs in the same group.

  To train baseline models:

  • Full positional encoding: omit the --barf_c2f argument.
  • No positional encoding: add --arch.posenc!.

  If you want to evaluate a checkpoint at a specific iteration number, use --resume=<ITER_NUMBER> instead of just --resume.

• Training the original NeRF

  If you want to train the reference NeRF models (assuming known camera poses):

  # Blender
  python3 train.py --group=<GROUP> --model=nerf --yaml=nerf_blender --name=<NAME> --data.scene=<SCENE>
  python3 evaluate.py --group=<GROUP> --model=nerf --yaml=nerf_blender --name=<NAME> --data.scene=<SCENE> --data.val_sub= --resume
  
  # LLFF
  python3 train.py --group=<GROUP> --model=nerf --yaml=nerf_llff --name=<NAME> --data.scene=<SCENE>
  python3 evaluate.py --group=<GROUP> --model=nerf --yaml=nerf_llff --name=<NAME> --data.scene=<SCENE> --resume

  If you wish to replicate the results from the original NeRF paper, use --yaml=nerf_blender_repr or --yaml=nerf_llff_repr instead for Blender or LLFF respectively. There are some differences, e.g. NDC will be used for the LLFF forward-facing dataset. (The reference NeRF models considered in the paper do not use NDC to parametrize the 3D points.)

• Planar image alignment experiment

  If you want to try the planar image alignment experiment, run:

  python3 train.py --group=<GROUP> --model=planar --yaml=planar --name=<NAME> --seed=3 --barf_c2f=[0,0.4]

  This will fit a neural image representation to a single image (default to data/cat.jpg), which takes a couple of minutes to optimize on a modern GPU. The seed number is set to reproduce the pre-generated warp perturbations in the paper. For the baseline methods, modify the arguments similarly as in the NeRF case above:

  • Full positional encoding: omit the --barf_c2f argument.
  • No positional encoding: add --arch.posenc!.

  A video vis.mp4 will also be created to visualize the optimization process.

• Visualizing the results

  We have included code to visualize the training over TensorBoard and Visdom. The TensorBoard events include the following:

  • SCALARS: the rendering losses and PSNR over the course of optimization. For BARF, the rotational/translational errors with respect to the given poses are also computed.
  • IMAGES: visualization of the RGB images and the RGB/depth rendering.

  We also provide visualization of 3D camera poses in Visdom. Run visdom -port 9000 to start the Visdom server.
  The Visdom host server is default to localhost; this can be overridden with --visdom.server (see options/base.yaml for details). If you want to disable Visdom visualization, add --visdom!.

  The extract_mesh.py script provides a simple way to extract the underlying 3D geometry using marching cubes. Run as follows:

  python3 extract_mesh.py --group=<GROUP> --model=barf --yaml=barf_blender --name=<NAME> --data.scene=<SCENE> --data.val_sub= --resume

  This works for both BARF and the original NeRF (by modifying the command line accordingly). This is currently supported only for the Blender dataset.

---

Codebase structure

The main engine and network architecture in model/barf.py inherit those from model/nerf.py. This codebase is structured so that it is easy to understand the actual parts BARF is extending from NeRF. It is also simple to build your exciting applications upon either BARF or NeRF -- just inherit them again! This is the same for dataset files (e.g. data/blender.py).

To understand the config and command lines, take the below command as an example:

python3 train.py --group=<GROUP> --model=barf --yaml=barf_blender --name=<NAME> --data.scene=<SCENE> --barf_c2f=[0.1,0.5]

This will run model/barf.py as the main engine with options/barf_blender.yaml as the main config file. Note that barf hierarchically inherits nerf (which inherits base), making the codebase customizable.
The complete configuration will be printed upon execution. To override specific options, add --<key>=value or --<key1>.<key2>=value (and so on) to the command line. The configuration will be loaded as the variable opt throughout the codebase.

Some tips on using and understanding the codebase:

• The computation graph for forward/backprop is stored in var throughout the codebase.
• The losses are stored in loss. To add a new loss function, just implement it in compute_loss() and add its weight to opt.loss_weight.<name>. It will automatically be added to the overall loss and logged to Tensorboard.
• If you are using a multi-GPU machine, you can add --gpu=<gpu_number> to specify which GPU to use. Multi-GPU training/evaluation is currently not supported.
• To resume from a previous checkpoint, add --resume=<ITER_NUMBER>, or just --resume to resume from the latest checkpoint.
• (to be continued....)

---

If you find our code useful for your research, please cite

@inproceedings{lin2021barf,
  title={BARF: Bundle-Adjusting Neural Radiance Fields},
  author={Lin, Chen-Hsuan and Ma, Wei-Chiu and Torralba, Antonio and Lucey, Simon},
  booktitle={IEEE International Conference on Computer Vision ({ICCV})},
  year={2021}
}

Please contact me (chlin@cmu.edu) if you have any questions!