The method takes low-light monocular images as input, and predicts frame-to-frame relative rotation and unit translation direction. To show trajectory, the translation direction is scaled by the ground-truth norm.
Visual odometry is a fundamental technology for autonomous systems, including robotics and self-driving cars. However, its performance often degrades in low-light conditions due to inherently low contrast, significant noise amplified by high gains, and motion blur induced by long exposure times. To mitigate these issues, we introduce a robust monocular visual odometry pipeline, named LOL-VO, that integrates a novel coarse-to-fine low-light enhancement network with a low-light robust learning-based pose estimator. The enhancement network employs histogram equalization on down-scaled images and progressively restores details through learnable gamma correction, achieving not only high-speed enhancement but also effective noise suppression. The pose network first fuses illumination-robust global features from the DINOv3 foundation model with local convolutional features, then estimates low-resolution optical flow through iterative refinement, and finally regresses the camera ego-motion using a differentiable weighted epipolar optimization block. Experiments on synthetic/real-world and indoor/outdoor scenarios show that LOL-VO achieves competitive VO accuracy and efficient visual enhancement across challenging low-light conditions.
git clone https://github.com/HITCSC/LOL-VO.git
cd ./LOL-VO
conda create -n lol python=3.10
conda activate lol
python -m pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple/
Move to this page, download the right one, and run
conda install pytorch3d-0.7.8-py310_cu121_pyt241.tar.bz2
(It seems acceptable with minor version mismatches.)
Download from here and put it under models/dinov3/checkpoints
Download our testing splits: TartanAir-v1
Run LOL-VO.
python -m predict --dataset tartanair
You can manually specify the running sequence in configs/tartanair.yaml.
The data is collected in our campus on December 2025. You can download from: self-collected dataset
The weight can generalize to the self-collected dataset without any finetuning.
python -m predict --dataset my_dataset
You can manually specify the running sequence in configs/my_dataset.yaml.
If you found this work to be useful in your own research, please considering citing the following information. [TODO]
This software is MIT licensed.
Copyright (c) [2026] Anonymous Authors
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.