1. Docker image is built for FAST-Calib, Kalibr, RealSense driver, Livox MID-360 SDK and allan_variance_ros.
4. Multi-scene joint calibration now searches all C(N,3) scene combinations and picks the trio with the lowest RMSE.
# ============================================================
# 1. Build the Docker image (on the host, from repo root)
# ============================================================
docker build -t kalibr_fastcalib_ros1_realsense:latest .# ============================================================
# 2. Allow the container to open GUI windows on the host's X server
# (run on the HOST, not inside the container)
# ============================================================
xhost +local:root# ============================================================
# 3. Start the container with GPU + X11 forwarding
# ============================================================
sudo docker run -it \
--env="DISPLAY=$DISPLAY" \
--volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" \
--gpus all \
kalibr_fastcalib_ros1_realsense:latest /bin/bash
# The ENTRYPOINT already sources /opt/ros/noetic/setup.bash and
# /catkin_ws/devel/setup.bash and drops you into /catkin_ws, so
# you do NOT need to run `source devel/setup.bash` manually.# ============================================================
# 4. (Optional) Rebuild fast_calib if you edited C++ sources
# The Dockerfile already builds everything, so skip this on
# a fresh container.
# ============================================================
cd /catkin_ws
catkin build fast_calib
source devel/setup.bash# ============================================================
# 5. Place your rosbags under calib_data/
# Each scene = one rosbag containing the LiDAR topic and
# the camera image topic. You need >= 3 scenes for the
# multi-scene joint calibration step.
# ============================================================
# e.g. /catkin_ws/src/FAST-Calib/calib_data/DataBag_2026-04-09-10-20-34/data.bag# ============================================================
# 6. Extract the first camera frame from each rosbag to an image.png
# IMPORTANT: edit BAG_PATH and TOPIC at the top of the script
# for EACH rosbag before running it.
# - RealSense: TOPIC = '/camera/color/image_raw/compressed'
# - Fisheye: TOPIC = '/right_camera/image/compressed'
# ============================================================
cd /catkin_ws/src/FAST-Calib # script uses ./calib_data/... (relative)
vim saveimg_rosbag.py # edit BAG_PATH and TOPIC
python3 saveimg_rosbag.py
# Expected output:
# Successfully saved compressed frame to ./calib_data/<BagName>/image.pngExamples of the distance-filter tool output:
-
📷 RealSense example:
-
🐟 Fisheye example:
# ============================================================
# 7. Pick a distance-filter box for each rosbag
# Open3D window opens:
# - Hold Shift + left-click to pick the 4 corners of the
# calibration board (>= 4 points required).
# - Press Q to close the window.
# The tool writes <pcd_name>.txt into the output dir with
# x_min/x_max/y_min/y_max/z_min/z_max values (board bbox
# expanded by 0.2 m on every side).
# ============================================================
cd /catkin_ws
python3 src/FAST-Calib/scripts/distance_filter_tool.py \
./src/FAST-Calib/calib_data/DataBag_2026-04-09-10-20-34/data.bag \
./src/FAST-Calib/output
# Example Open3D log:
# Processing: livox_CustomMsg_inten_ascii.pcd
# [Open3D INFO] Picked point #10815894 (1.8, 0.69, 1.6)
# [Open3D INFO] Picked point # 5603795 (1.8, -0.62, 1.6)
# [Open3D INFO] Picked point # 1808882 (2.1, -0.66, 0.78)
# [Open3D INFO] Picked point # 566301 (2.1, 0.62, 0.78)
# Then copy x_min/x_max/y_min/y_max/z_min/z_max from the generated
# .txt into the Distance filter block of qr_params.yaml (step 8).Examples of the distance-filter tool output:
-
📷 RealSense example:
-
🐟 Fisheye example:
# ============================================================
# 8. Edit qr_params.yaml for THIS rosbag
# Set per-sensor (edit once per sensor suite):
# - fx, fy, cx, cy, k1, k2, p1, p2 (camera intrinsics)
# - marker_size, delta_width/height_qr_center,
# delta_width/height_circles, circle_radius
# - lidar_topic (e.g. /livox/lidar, /ouster/points, /hesai/pandar)
# Set per-rosbag (edit for every scene):
# - bag_path, image_path
# - x_min/x_max/y_min/y_max/z_min/z_max (from step 7)
# To disable the distance filter, comment out the 6 x/y/z
# lines and the code will use all points.
# ============================================================
vim src/FAST-Calib/config/qr_params.yaml# ============================================================
# 9. Run single-scene calibration for this rosbag
# RViz will open; the node writes its per-scene result into
# the output_path configured in qr_params.yaml.
# Repeat steps 6-9 for every rosbag (>= 3 total).
# ============================================================
roslaunch fast_calib calib.launchExamples of the distance-filter tool output:
-
📷 RealSense example:
-
🐟 Fisheye example:
# ============================================================
# 10. After at least 3 scenes have been processed, run the
# multi-scene joint calibration. It searches all C(N,3)
# scene combinations and picks the trio with the lowest
# RMSE (see summary item 4 above).
# ============================================================
roslaunch fast_calib multi_calib.launchExample of RealSense D435 & Livox MID-360 lidar:
[Result] RMSE: 0.0036 m
[Result] Multi-scene calibration: extrinsic parameters T_cam_lidar =
-0.001555 -0.999993 0.003546 0.027207
0.443624 -0.003868 -0.896205 -0.224154
0.896212 0.000179 0.443627 -0.026632
0.000000 0.000000 0.000000 1.000000Example of Fisheye Camera & Livox MID-360 lidar:
[Result] RMSE: 0.0032 m
[Result] Multi-scene calibration: extrinsic parameters T_cam_lidar =
-0.007156 -0.999971 0.002528 -0.050784
0.447482 -0.005463 -0.894276 -0.129908
0.894264 -0.005269 0.447509 -0.006086
0.000000 0.000000 0.000000 1.000000FAST-Calib is an efficient target-based extrinsic calibration tool for LiDAR-camera systems (eg., FAST-LIVO2).
Key highlights include:
- Support solid-state and mechanical LiDAR.
- No need for any initial extrinsic parameters.
- Achieve highly accurate calibration results in just one seconds.
In short, it makes extrinsic calibration as simple as intrinsic calibration.
Related paper:
FAST-Calib: LiDAR-Camera Extrinsic Calibration in One Second
📬 For further assistance or inquiries, please feel free to contact Chunran Zheng at zhengcr@connect.hku.hk.
Left: Example of Mid360 LiDAR calibration. Right: Point cloud colored with the calibrated extrinsics.
Circular hole extraction supports multiple LiDAR models.
PCL>=1.8, OpenCV>=4.0.
- Prepare the static acquisition data in the
calib_datafolder (see Single-scene Calibration Sample Data from Mid360, Avia and Ouster, and Multi-scene Calibration Sample Data from Avia):
- rosbag containing point cloud messages
- corresponding image
- Run the single-scene calibration process:
roslaunch fast_calib calib.launch- After completing Step 2 for at least three different scenes, you can perform multi-scene joint calibration:
roslaunch fast_calib multi_calib.launch- Customize the calibration target in the image below, with the CAD model available here.
- Collect data from three scenes, with placement illustrated below, and record them into the corresponding rosbags.
- Provide the instrinsic matrix in
qr_params.yaml. - Set distance filter in
qr_params.yamlfor board point cloud (extra points are acceptable). - Calibrate now!
💡 Note: You can run scripts/distance_filter_tool.py to quickly obtain suitable filter parameters.
Left: Actual calibration target | Right: Technical drawing with annotated dimensions.
Placement of the calibration target for multi-scene data collection: (a) facing forward, (b) oriented to the right, (c) oriented to the left.
The calibration target design is based on the velo2cam_calibration.
For further details on the algorithm workflow, see this document.
Special thanks to Jiaming Xu for his support, Haotian Li for the equipment, and the velo2cam_calibration algorithm.