Parallel Inversion NeRF

Overview

This repository is the official implementation of the paper Parallel Inversion of Neural Radiance Fields for Robust Pose Estimation by Lin et al., ICRA 2023 (full citation below). For videos, please visit the Parallel Inversion NeRF project site.

In this work, we present a parallelized optimization method based on fast Neural Radiance Fields (NeRF) for estimating 6-DoF pose of a camera with respect to an object or scene. Given a single observed RGB image of the target, we can predict the translation and rotation of the camera by minimizing the residual between pixels rendered from a fast NeRF model and pixels in the observed image. We integrate a momentum-based camera extrinsic optimization procedure into Instant Neural Graphics Primitives, a recent exceptionally fast NeRF implementation. By introducing parallel Monte Carlo sampling into the pose estimation task, our method overcomes local minima and improves efficiency in a more extensive search space. We also show the importance of adopting a more robust pixel-based loss function to reduce error. Experiments demonstrate that our method can achieve improved generalization and robustness on both synthetic and real-world benchmarks. The algorithm takes an average of 25 seconds on an NVIDIA GTX 3090 GPU.

Installation

The code was tested on Ubuntu 20.04, with Anaconda Python 3.9. Other versions should be possible with some accuracy difference. NVIDIA GPUs are needed for both training and testing.

1. Preparation for Instant NGP

$ sudo apt-get install build-essential git python3-dev python3-pip libopenexr-dev libxi-dev \
libglfw3-dev libglew-dev libomp-dev libxinerama-dev libxcursor-dev

2. Clone this repo:

$ REPO_ROOT=/path/to/clone/ParallelInversion
$ git clone --recursive https://github.com/NVlabs/ParallelInversion.git $REPO_ROOT

3. Create an Anaconda environment or create your own virtual environment

$ conda create -n pinerf python=3.9
$ conda activate pinerf
$ pip install -r requirements.txt

4. Build the project using CMake

$ cmake . -B build -DCMAKE_BUILD_TYPE=RelWithDebInfo $ cmake --build build --config RelWithDebInfo -j 

Dataset

To play with scenes presented in the paper, download the data here. Place the downloaded dataset according to the following directory structure:

 ├── data/nerf │ ├── nerf_llff_data │ │ └── fern │ │ └── fortress | | └── ...
| ├── nerf_synthetic
| | └── chair
| | └── drums | | └── ...

Following NeRF-Pytorch and Instant-NGP, we further preprocess LLFF dataset by converting it to the NeRF format. We provide the script colmap2nerf.py to convert the specific dataset format to the NeRF format (saving a json file).

$ cd REPO_ROOT/scripts
$ export PYTHONPATH='../'
$ python preprocess_llff.py

For a sanity check, you can visualize the converted dataset by running the following command:

$ cd REPO_ROOT/scripts
$ export PYTHONPATH='../'
$ python run_debug.py --config config/debug/default.yaml

Evaluation

We evaluate our method on NeRF Synthetic and LLFF datasets.

$ cd REPO_ROOT/scripts
$ export PYTHONPATH='../'## Main# Evaluation on NeRF Synthetic dataset
$ python exp_nerf_synthetic.py --config config/exp/exp_nerf_synthetic.yaml
$ python exp_nerf_synthetic.py --config config/exp/exp_nerf_synthetic_multiGuess.yaml
# Evaluation on LLFF dataset
$ python exp_nerf_llff.py --config config/exp/exp_nerf_llff.yaml
$ python exp_nerf_llff.py --config config/exp/exp_nerf_llff_multiGuess.yaml
## Ablation study# Evaluation on NeRF Synthetic dataset with simulated noise
$ python exp_nerf_synthetic.py --config config/exp/exp_nerf_synthetic_noisy.yaml --eval_mode ablation_rgb_loss
# Evaluation on LLFF dataset
$ python exp_nerf_llff.py --config config/exp/exp_nerf_llff.yaml --eval_mode ablation_rgb_loss

We also provide the options to generate visualization results. An example is shown below:

$ cd REPO_ROOT/scripts
$ export PYTHONPATH='../'
$ python run_pose_refinement.py --config config/demo/nerf_synthetic_hotdog_multiGuess.yaml

You can get the following result:

Acknowledgements

In this repository, we have used codes or datasets from the following repositories. We thank all the authors for sharing great codes or datasets.

• Instant-NGP
• NeRF-pytorch

Citation

Please cite the following if you use this repository in your publications:

@inproceedings{lin2023icra:parallel,
title={Parallel Inversion of Neural Radiance Fields for Robust Pose Estimation},
author={Lin, Yunzhi and Müller, Thomas and Tremblay, Jonathan and Wen, Bowen and Tyree, Stephen and Evans, Alex and Vela, Patricio A. and Birchfield, Stan},
booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
year={2023}
}

Licence

Parallel Inversion NeRF is licensed under the NVIDIA Source Code License - Non-commercial.