jike5 / LATITUDE

LATITUDE(ICRA2023)

This is the official repository of LATITUDE: Robotic Global Localization with Truncated Dynamic Low-pass Filter in City-scale NeRF.

Introduction

In this work, we present LATITUDE: Global Localization with Truncated Dynamic Low-pass Filter, which introduces a two-stage localization mechanism in large-scale NeRF. In place recognition stage, we train a regressor through images generated from trained NeRFs, which provides an initial value for global localization. In pose optimization stage, we minimize the residual between the observed image and rendered image by directly optimizing the pose on the tangent plane. To avoid falling into local optimum, we introduce a Truncated Dynamic Low-pass Filter (TDLF) for coarse-to-fine pose registration. We evaluate our method on both synthetic and real-world data and show its potential applications for high-precision navigation in large-scale city scenes.

Overview

This preliminary codebase currently only shows how to apply LATITUDE with mega-NeRF. However, it can work with the original NeRF as well.

Environment setup

Create a conda virtual environment and activate it.

conda env create -f environment.yml
conda activate LATITUDE

Prepare Dataset

• Mill 19

  • Mill19 Link

• UMAD

  • 百度云
  • Google Drive

Training

The following will describe how to use Mega-nerf for training, you can refer to cmusatyalab/mega-nerf for detailed usage instructions. You must use the code of this repository, for we make partial modifications to mega-nerf. We will use the UMAD dataset as an example, assuming that you download the UMAD dataset at the following location:

datasets/
|-- high-simple($DATA_DIR)
|   |-- train
|       |-- metadata
|       |-- rgbs
|   |-- train
|       |-- metadata
|       |-- rgbs
|   |-- coordinates.pt
|   |-- mappings.txt
|-- high-hard
|-- ...
|
LATITUDE/ 
|-- inerf
|-- mega_nerf
|-- mega-nerf-output # files during training 
|-- test
|-- ...

1) Generate the training partitions for each submodule

python create_cluster_masks.py --config config/${DATASET_NAME}.yaml --dataset_path $DATASET_PATH  --output $MASK_PATH --grid_dim $GRID_X $GRID_Y

Note: In ${DATASET_NAME}.yml, you should set ray_altitude_range according to your UAV height.

One example is:

python create_cluster_masks.py --config config/high_simple.yaml --dataset_path ../dataset/high-simple --output mega-nerf-output/mask/high_simple --grid_dim 1 1

2) Train each submodule

python train_mega_nerf.py --config_file config/${DATASET_NAME}.yml --exp_name $EXP_PATH/$EXP_PREFIX-${SUBMODULE_INDEX} --dataset_path $DATASET_PATH --chunk_paths $SCRATCH_PATH --cluster_mask_path ${MASK_PATH}/${SUBMODULE_INDEX}

Note: If there are multiple submodules, train each submodule separately, means change the value of ${SUBMODULE_INDEX}. The last folder name of $EXP_PATH needs to be the same as $EXP_PREFIX.

One example is:

python train_mega_nerf.py --config_file config/high_simple.yaml --exp_name mega-nerf-output/exp/high_simple/high_simple-0  --dataset_path ../dataset/high-simple --chunk_paths mega-nerf-output/chunk_dir --cluster_mask_path mega-nerf-output/mask/high_simple/0

3) Merge the trained submodules into a unified Mega-NeRF model

python merge_submodules.py --config_file config/${DATASET_NAME}.yaml  --ckpt_prefix ${EXP_PREFIX}- --centroid_path ${MASK_PATH}/params.pt --output $MERGED_OUTPUT

Note: Specify the -ckpt_prefix parameter as: ${EXP_PREFIX}-, please note that there is a short horizontal bar after it.

One example is:

python merge_submodules.py --config_file config/high_simple.yaml --ckpt_prefix mega-nerf-output/exp/high_simple/high_simple- --centroid_path mega-nerf-output/mask/high_simple/params.pt --output mega-nerf-output/merged/high_simple.pt --ckpt_iteration 150000

Pose Regressor

Change settings in config/config-posenet.txt.

python pose_regressor/script/run_posenet.py --config config/config_posenet.txt

Pose Optimization

Run the command to start the optimization, you can set the relevant noise parameters in the file config

python pose_optimization/run.py --data_dir $DATASET_PATH --nerf_config config/${DATASET_NAME}.yml --container_path $MERGED_OUTPUT

You can use --pose_regressor_input option to use PoseRegressor module's output as intial value.

Bibtex

If this work is helpful for your research, please cite the following BibTeX entry.

@article{zhu2023latitude,
  title={LATITUDE: Robotic Global Localization with Truncated Dynamic Low-pass Filter in City-scale NeRF},
  author={Zhu, Zhenxin and Chen, Yuantao and Wu, Zirui and Hou, Chao and Shi, Yongliang and Li, Chuxuan and Li, Pengfei and Zhao, Hao and Zhou, Guyue},
  booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={8326--8332},
  year={2023},
  organization={IEEE}
}

Contact

If you have any questions, please open a github issue or contact us.

Zhenxin Zhu: [e-mail], Yuantao Chen: [e-mail], Zirui Wu: [e-mail], Chao Hou: [e-mail]

Acknowledgements

We would like to thank mega-nerf, barf and dfnet for making their project public.