Multi-camera setups find widespread use across various applications, such as au- tonomous driving, as they greatly expand sensing capabilities. Despite the fast development of Neural radiance field (NeRF) techniques and their wide applica- tions in both indoor and outdoor scenes, applying NeRF to multi-camera systems remains very challenging. This is primarily due to the inherent under-calibration issues in multi-camera setup, including inconsistent imaging effects stemming from separately calibrated image signal processing units in diverse cameras, and system errors arising from mechanical vibrations during driving that affect rela- tive camera poses. In this paper, we present UC-NeRF, a novel method tailored for novel view synthesis in under-calibrated multi-view camera systems. Firstly, we propose a layer-based color correction to rectify the color inconsistency in different image regions. Second, we propose virtual warping to generate more viewpoint-diverse but color-consistent virtual views for color correction and 3D recovery. Finally, a spatiotemporally constrained pose refinement is designed for more robust and accurate pose calibration in multi-camera systems. Our method not only achieves state-of-the-art performance of novel view synthesis in multi- camera setups, but also effectively facilitates depth estimation in large-scale out- door scenes with the synthesized novel views.
# Clone the repo.
git clone https://github.com/kcheng1021/UC-NeRF.git
cd nerf
# Make a conda environment.
conda create --name ucnerf python=3.9
conda activate ucnerf
# Install requirements.
pip install -r requirements.txt
# Install other cuda extensions
pip install ./gridencoder
# Install nvdiffrast
git clone https://github.com/NVlabs/nvdiffrast
pip install ./nvdiffrast
# Install a specific cuda version of torch_scatter
# see more detail at https://github.com/rusty1s/pytorch_scatter
CUDA=cu113
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.12.0+${CUDA}.html
Here we present a sample Waymo scene in Waymo-100613.
The detail tutorial of the preprocessing of Waymo dataset can be refered in StreetSurf.
# getting the refined poses
cd pose_refinement
# following the instruction in the submodule
# getting the depth for virtual warping
cd mvs
# following the instruction in the submodule
cd nerf
# including the bash scripts of train, eval nerf model, and render video
bash scripts/train_waymo.sh
For non-commercial use, this code is released under the LICENSE. For commercial use, please contact Xuejin Chen.
If you find this project useful in your research, please consider cite:
@inproceedings{cheng2023uc,
title={UC-NERF: Neural Radiance Field for under-calibrated multi-view cameras},
author={Cheng, Kai and Long, Xiaoxiao and Yin, Wei and Wang, Jin and Wu, Zhiqiang and Ma, Yuexin and Wang, Kaixuan and Chen, Xiaozhi and Chen, Xuejin},
booktitle={The Twelfth International Conference on Learning Representations},
year={2023}
}
Thanks to Zip-NeRF-Pytorch for amazing zipnerf pytorch implementation. Thanks to COLMAP and CER-MVS for providing the foundation for our work.