Global-LVBA is a globally consistent LiDAR–Visual Bundle Adjustment system for refinement after LiDAR-inertial-visual odometry (e.g., FAST-LIVO2).
If you want to push accuracy to the next level, Global-LVBA further optimizes:

• point-cloud map consistency (reducing layering/stratification artifacts).
• camera poses towards pixel-level reprojection accuracy.

📬 For further assistance or inquiries, please feel free to contact Chunran Zheng at zhengcr@connect.hku.hk.

Global-LVBA Cover
This figure shows qualitative comparisons before and after Global-LVBA refinement from both texture and geometric-structure perspectives. The initial poses are provided by FAST-LIVO2.

• Python 3.8
• Ceres Solver 2.1.0
• Eigen 3.3.7
• OpenCV 4.2.0

cd catkin_ws/src/Global-LVBA
git submodule update --init --recursive
cd SiftGPU && mkdir build && cd build && cmake .. && make
cd catkin_ws && catkin_make

Global-LVBA/
└── dataset/
 └── your_sequence_name/
 ├── all_image/
 │ ├── 1661398632.022152.png # image named by timestamp
 │ ├── 1661398632.121881.png
 │ ├── ...
 │ └── image_poses.txt # camera poses (timestamp-aligned)
 ├── all_pcd_body/
 │ ├── 1661398632.022152.pcd # point cloud named by timestamp
 │ ├── 1661398632.121881.pcd
 │ ├── ...
 │ └── lidar_poses.txt # LiDAR poses (timestamp-aligned)
 ├── colmap/
 │ ├── match.db # feature matching database
 │ └── sparse/ # optional: COLMAP output for 3DGS
 │ └── 0/
 │ ├── images.txt
 │ └── points3D.txt
 ├── depth/ # optional: generated depth maps
 └── reproj/ # optional: reprojection error logs

The LVBA-Dataset (Google Drive) is generated from the raw rosbag in FAST-LIVO2-Dataset (Google Drive).

💡 Note: To prepare your own data with FAST-LIVO2, set pcd_save_en: false, type: 1, and img_save_en: true in the FAST-LIVO2 config to directly generate the required files for Global-LVBA.

Download the example data from LVBA-Dataset.

roslaunch Global-LVBA lvba.launch

💡 Note: We recommend using the .db file generated by COLMAP to obtain feature correspondences. The built-in SiftGPU extraction/matching in this repo is provided for convenience, but its matching quality is generally inferior to COLMAP.

The source code of this package is released under the MIT license.