We take the images in examples/thin_catbus for example. These images were captured with an ArUco board. The pattern image of this board can be found in ./static/aruco_board.png.

Step 1. Build the calibration code (c++)

Dependencies:

• cnpy (https://github.com/rogersce/cnpy)
• OpenCV

Run commands

cd aruco_preprocess
mkdir build
cd build
cmake .. & make
cd ..

Step 2. Get the preprocessed data

First indicate data_dir,n_images in run.sh. For example, data_dir=./example/thin_catbus

Run commands

bash run.sh

Then the preprocessed data can be found in ${data_dir}/preprocessed. Here we define the region of interest as the bounding sphere of the calibration board.

Step 1. Run COLMAP SfM

Run commands

cd colmap_preprocess
python img2poses.py ${data_dir}

After running the commands above, a sparse point cloud is saved in ${data_dir}/sparse_points.ply.

Step 2. Define the region of interest

The raw sparse point cloud may be noisy and may not be appropriate to define a region of interest (The white frame indicates the bounding box of the point cloud):

And you may need to clean it by yourself (here we use Meshlab to clean it manually). After cleaning:

Save it as ${data_dir}/sparse_points_interest.ply.

Then run the commands:

python gen_cameras.py ${data_dir}

Then the preprocessed data can be found in ${data_dir}/preprocessed.

Here we just use the image without undistortion in the second option. To get better results, you may need to undistort your images in advance.

The python scripts to run COLMAP SfM are heavily borrowed from LLFF: https://github.com/Fyusion/LLFF