CalibAnything

This package provides an automatic and target-less LiDAR-camera extrinsic calibration method using Segment Anything Model. The related paper is Calib-Anything: Zero-training LiDAR-Camera Extrinsic Calibration Method Using Segment Anything. For more calibration codes, please refer to the link SensorsCalibration.

Prerequisites

• pcl 1.10
• opencv
• eigen 3

Compile

git clone https://github.com/OpenCalib/CalibAnything.git
cd CalibAnything
# mkdir build
mkdir -p build && cd build
# build
cmake .. && make

Run Example

We provide examples of two dataset. You can download the processed data at Google Drive or BaiduNetDisk:

# baidunetdisk
Link: https://pan.baidu.com/s/1qAt7nYw5hYoJ1qrH0JosaQ?pwd=417d 
Code: 417d

Run the command:

cd CalibAnything
./bin/run_lidar2camera ./data/kitti/calib.json # kitti dataset
./bin/run_lidar2camera ./data/nuscenes/calib.json # nuscenes dataset

Test your own data

Data collection

• Several pairs of time synchronized RGB images and LiDAR point cloud (intensity is needed). One pair of data can also be used to calibrate, but the results may be ubstable.
• The intrinsic of the camera and the initial guess of the extrinsic.

Preprocessing

Generate masks

Follow the instructions in Segment Anything and generate masks of your image.

1. First download a model checkpoint. You can choose vit-l.

2. Install SAM

# environment: python>=3.8, pytorch>=1.7, torchvision>=0.8

git clone git@github.com:facebookresearch/segment-anything.git
cd segment-anything; pip install -e .
pip install opencv-python pycocotools matplotlib onnxruntime onnx

3. Run

python scripts/amg.py --checkpoint <path/to/checkpoint> --model-type <model_type> --input <image_or_folder> --output <path/to/output>

# example(recommend parameter)
python scripts/amg.py --checkpoint sam_vit_l_0b3195.pth --model-type vit_l --input ./data/kitti/000000/images/  --output ./data/kitti/000000/masks/ --stability-score-thresh 0.9 --box-nms-thresh 0.5 --stability-score-offset 0.9

Data folder

The hierarchy of your folders should be formed as:

YOUR_DATA_FOLDER
├─calib.json
├─pc
|   ├─000000.pcd
|   ├─000001.pcd
|   ├─...
├─images
|   ├─000000.png
|   ├─000001.png
|   ├─...
├─masks
|   ├─000000
|   |   ├─000.png
|   |   ├─001.png
|   |   ├─...
|   ├─000001
|   ├─...

Processed masks

For large masks, we only use part of it near the edge.

python processed_mask.py -i <YOUR_DATA_FOLDER>/masks/ -o <YOUR_DATA_FOLDER>/processed_masks/

Edit the json file

Content description

• cam_K: camera intrinsic matrix
• cam_dist: camera distortion coefficient. [k1, k2, p1, p2, p3, ...], use the same order as opencv
• T_lidar_to_cam: initial guess of the extrinsic
• T_lidar_to_cam_gt: ground-truth of the extrinsic (Used to calculate error. If not provided, set "available" to false)
• img_folder: the path to images
• mask_folder: the path to masks
• pc_folder: the path to point cloud
• img_format: the suffix of the image
• pc_format: the suffix of the point cloud (support pcd or kitti bin)
• file_name: the name of the input images and point cloud
• min_plane_point_num: the minimum number of point in plane extraction
• cluster_tolerance: the spatial cluster tolerance in euclidean cluster (set larger if the point cloud is sparse, such as the 32-beam LiDAR)
• search_num: the number of search times
• search_range: the search range for rotation and translation
• point_range: the approximate height range of the point cloud projected onto the image (the top of the image is 0.0 and the bottom of the image is 1.0)
• down_sample: the point cloud downsample voxel size (if don't need downsample, set the "is_valid" to false)
• thread: the number of thread to reduce calibration time

Calibration

./bin/run_lidar2camera <path-to-json-file>

Output

• initial projection: init_proj.png, init_proj_seg.png
• gt projection: gt_proj.png, gt_proj_seg.png
• refined projection: refined_proj.png, refined_proj_seg.png
• refined extrinsic: extrinsic.txt

Citation

If you find this project useful in your research, please consider cite:

@misc{luo2023calibanything,
      title={Calib-Anything: Zero-training LiDAR-Camera Extrinsic Calibration Method Using Segment Anything}, 
      author={Zhaotong Luo and Guohang Yan and Yikang Li},
      year={2023},
      eprint={2306.02656},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}