Complementary-pathway Spatial-enhanced Visual Odometry

IROS25_0435_VI_fi.mp4

This repository contains the source code for our paper:

@article{Lin2025complementary,
  title={Complementary-pathway Spatial-enhanced Visual Odometry for Extreme Environments with Brain-inspired Vision Sensors},
  author={Yihan, Lin and Zhaoxi, Zhang and Taoyi, Wang and Yuguo, Chen and Rong, Zhao},
  journal={International Conference on Intelligent Robots and Systems (IROS)},
  year={2025}
}

Setup and Installation

Clone the repo

git clone https://github.com/JesseZZZZZ/CSVO.git
cd CSVO

Create and activate the csvo anaconda environment

conda env create -f environment.yml
conda install cuda -c nvidia/label/cuda-11.3.1
conda activate csvo

Next install the packages

wget https://gitlab.com/libeigen/eigen/-/archive/3.4.0/eigen-3.4.0.zip
unzip eigen-3.4.0.zip -d thirdparty

# install CSVO
pip install .

# download models and data (~2GB)
./download_models_and_data.sh

Dataset Preprocessing

TartanAir

You can refer to https://github.com/castacks/tartanair_tools for the TartanAir dataset. We suggest you to arrange the dataset as follows:

├── datasets
    ├── TartanAir.pickle
    ├── TartanAir
        ├── abandonedfactory
        ├── abandonedfactory_night
        ├── ...
        ├── westerndesert
    ...

TartanAirAug

We provide simple augmentation methods for TartanAir. You can run the following command to generate the augmented dataset:

python augmentation/augmentation.py /path/to/TartanAir

Tianmouc data

We currently only provide the Tianmouc-VO dataset, which is already preprocessed. You can download it from here. If you want to record some Tianmouc data with your own Tianmouc camera, you can refer to here, where we provide the decoding and alignment code for Tianmouc camera.

Training

To train the model, you can run the following command:

python train.py --steps=240000 --lr=0.00008 --name=[dataset name]  --training_type  [input mode]

The training process requires about 5 days on a single NVIDIA RTX 3090 GPU. Example training logs can be seen at examples

Evaluation

We provide evaluation scripts for TartanAir, TianmoucVO, and Apollo. Up to date result logs on these datasets can be found in the logs directory.

you can use sh scripts under ./scripts flder for convenience

TartanAir:

Results on the validation split and test set can be obtained with the command:

python evaluate_tartan.py --trials=5 --split=validation --plot --save_trajectory

TartanAirAug (Augmented data of TartanAir):

To run CSVO and our baseline (DPVO), you can directly run these scripts:

bash scrips/evaluate_tartan_augmented_csvo.sh
bash scrips/evaluate_tartan_augmented_dpvo.sh

Tianmouc-VO:

To run CSVO and our baseline (DPVO), you can directly run these scripts:
bash scrips/evaluate_Tianmouc_csvo.sh
bash scrips/evaluate_Tianmouc_dpvo.sh

Apollo:

python evaluate_apollo.py --trials=5 --plot --save_trajectory

tmdat data inference support

please install tianmoucv first:

pip install tianmoucv

then try our demo:

bash ./scripts/run_demo.sh

Acknowledgements

• This repository is built on DPVO.
• If you have problems on installation, you can refer to DPVO's Issues, which has a similar conda environment as ours.