There exist many resources for Camera Calibration and this has become a standard operating procedure. Here, we use OpenCV to calibrate single or stereo cameras using chessboard or ChArUco boards. We emphasize the importance of high-quality calibration boards that are on flat surfaces and not warped.
Alternative calibration options include importing calibration data produced with the Calibrator calib.io Software, which relies on a different matrix-optimization approach and other flexible parameters. A python-based conversion code is available that converts output from Calibrator to the OpenCV XML camera matrix format.
Table of Contents
- Usage
- Best practices for a good calibration
- Our testing
- Visualizing and comparing camera calibrations
- Visualize up to 4 distortion patterns from camera calibration filesVisualize-up-to-4-distortion-patterns-from-camera-calibration-file)
- compare the distortion fields of two camera calibration files
The parameters have been optimized for the cameras that are available at the Geological Remote Sensing lab at the University of Potsdam. These are several Sony alpha-6 (24 MP), Sony alpha-7 (40 MP), and Fuji X-100 (24 MP) all with 55 mm or 85 mm fixed lenses.
The python code python/single_camera_calibration_charuco_chess_openCV.py can be called from the command line. Use single_camera_calibration_charuco_chess_openCV.py -h to obtain a short help and description of the parameters.
The code will read all calibration images from one directory, plots a summary figure showing all photos, performs the camera calibration, and writes the distortion coefficient and intrinsic camera calibration to an OpenCV XML file.
Example of 49 photos showing a chessboard pattern for camera calibration (Sony alpha-7 55 mm lense):
All detected chessboard intersection - make sure that points are also taken from the corners of the image (Sony alpha-7 55 mm lense):
Example camera calibration and pixel distortion using intrinsic and distorted parameters (Sony alpha-7 55 mm lense):
Example call from Ubuntu command line (expecting OpenCV to be installed).
camA_initial_CC='cam_A_calib_9parameters_fine_charuco_20Feb2022.xml'
charuco_ifiles_camA='sony_stereo_f13_iso1600/charuco/black_a_stereo/DSC*.JPG'
camA_charuco_savexml_file='sony_stereo_f13_iso1600/charuco/cam_A_black_calib_9parameters_fine_charuco_25Mar2022.xml'
camA_CC_comparison_3panel_png='sony_stereo_f13_iso1600/charuco/CC_comparison_3panel.png'
camA_CC_comparison_1panel_png='sony_stereo_f13_iso1600/charuco/CC_comparison_1panel.png'
camA_Height=4000
camA_Width=6000
focal_length_pixels=9000
single_camera_calibration_charuco_chess_openCV.py --camA_initial_CC $camA_initial_CC \
--charuco_ifiles_camA $charuco_ifiles_camA \
--camA_charuco_savexml_file $camA_charuco_savexml_file \
--camA_CC_comparison_3panel_png $camA_CC_comparison_3panel_png \
--camA_CC_comparison_1panel_png $camA_CC_comparison_1panel_png \
--camA_Height $camA_Height --camA_Width $camA_Width \
--focal_length_pixels $focal_length_pixelsUsing no initial calibration file and this requires setting the parameters camA_Height, camA_Width, and focal_length_pixels.
chess_ifiles_camA='near/DSC*.JPG'
camA_chess_savexml_file='sony_alpha7_55m_CC_05July2022.xml'
camA_chess_75pbest_savexml_file='sony_alpha7_55m_CC_75p_05July2022.xml'
camA_CC_comparison_3panel_png='sony_alpha7_55m_CC_chess_comparison_3panel.png'
camA_CC_comparison_1panel_png='sony_alpha7_55m_CC_chess_comparison_1panel.png'
camA_Height=5304
camA_Width=7952
focal_length_pixels=12675
single_camera_calibration_charuco_chess_openCV.py \
--chess_ifiles_camA $chess_ifiles_camA \
--camA_chess_savexml_file $camA_chess_savexml_file \
--camA_chess_75pbest_savexml_file $camA_chess_75pbest_savexml_file \
--camA_CC_comparison_3panel_png $camA_CC_comparison_3panel_png \
--camA_CC_comparison_1panel_png $camA_CC_comparison_1panel_png \
--camA_Height $camA_Height --camA_Width $camA_Width \
--focal_length_pixels $focal_length_pixelsUsing no initial calibration file and this requires setting parameters for camera calibration.
chess_ifiles_camA='near/DSC*.JPG'
camA_chess_savexml_file='sony_alpha7_85m_CC_05July2022.xml'
camA_chess_75pbest_savexml_file='sony_alpha7_85m_CC_75p_05July2022.xml'
camA_CC_comparison_3panel_png='sony_alpha7_85m_CC_chess_comparison_3panel.png'
camA_CC_comparison_1panel_png='sony_alpha7_85m_CC_chess_comparison_1panel.png'
camA_Height=5304
camA_Width=7952
focal_length_pixels=18918
single_camera_calibration_charuco_chess_openCV.py \
--chess_ifiles_camA $chess_ifiles_camA \
--camA_chess_savexml_file $camA_chess_savexml_file \
--camA_chess_75pbest_savexml_file $camA_chess_75pbest_savexml_file \
--camA_CC_comparison_3panel_png $camA_CC_comparison_3panel_png \
--camA_CC_comparison_1panel_png $camA_CC_comparison_1panel_png \
--camA_Height $camA_Height --camA_Width $camA_Width \
--focal_length_pixels $focal_length_pixelsIf you have an exported JSON calibration file from a calibration done using Calib.io, you convert it using our conversion script. If json_dir or xml_dir is not supplied, the script will look for JSON files in the current directory and ouput the related XML files in a newly created xml/ directory.
python python/calib-to-opencv.py --json_dir="path/to/json" --xml_dir="path/to/output"For help use:
python python/calib-to-opencv.py -hBefore capturing calibration images it is critical to ensure the camera is configured and mounted properly, the board is of high quality, and the scene is set appropriately to ensure a smooth calibration process.
Calibration works best when the camera's internal settings (e.g. f-stop, shutter speed, ISO) are fixed and constant for the calibration session. The cameras used here were placed into "Manual" mode (a feature on the majority of high-quality digital cameras today) and f-stop, shutter speed, and ISO manually set according to the lighting conditions at the time of the sessions.
When taking the photos it is important to minimize blur, so if it is not possible to use a remote shutter when capturing the images, consider enabling a short self-timer to avoid vibration at the time of capture.
The Sony cameras used in this study were also equipped with a "SteadyShot" feature which was disabled prior to shooting because, when enabled, it results in small adjustments being made to the internal mechanics of the lens to physically eliminate motion blur, but at the cost of slightly altering the lens's internal parameters and resulting in poorer calibration results. It is recommended that any automatic blur-reduction features are disabled.
The camera should be mounted on a tripod to ensure that its internal and geometric parameters remain constant throughout the session. This is crucial for a good calibration and handheld photos result in a significantly worse result.
We recommend the use of a high-quality calibration board such as the aluminum composite checkerboard or CharuCo targets made by Calib.io, because any inconsistencies in the target's flatness or pattern will affect the calibration results. Target stiffness is also important as it is recommended to place the board in multiple orientations (often requiring it to be leaned at an angle against another object).
We found that the checkerboard targets tended to provide better calibration results, but because the entire checkerboard needs to be within the frame at all times it can be somewhat tricky to align it perfectly in the corners (where the most distortion often is found). The CharuCo targets help with this issue as they do not need to be entirely in the frame, but at the cost of slightly worse results.
|
|
|
|
|
If performing the calibration indoors, a well-lit environment with minimal shadows is key. Outdoor environments during noon times will work well, but avoid glare or reflections from the boards. LED lighting panels are good for this. Place the tripod with the camera in front of the target area and tilt the camera to a slightly oblique orientation to the ground. In fact, we found that calibrations conducted with the camera at an angle to the ground as opposed to looking straight down resulted in lower calibration error. For quick and consistent target placement during the shooting, we found it helped to first place tape along the edge of the camera frame. This can be done by placing the target in one corner of the frame as precisely as possible and then using it as a guide to place the tape. Repeat for the other three corners.
We emphasize that a fixed camera setup is not required for a successful calibration, but it often helps having a tripod-mounted camera to avoid blur. In recent work we show that a moving camera with fixed calibration board results in the similar parameters as a fixed camera (with moving checkerboard).
Once you have the frame marked out, we recommend taking at least 20-30 photos (better 100) with the target covering every part of the frame at least once. Placing the target at oblique angles to the camera is also recommended, and for this we simply leaned the target up against an item in the lab and attempted to capture it in several different areas of the frame. If using the checkerboard, we had best results when placing it as close to the edge and in the corners of the frame as possible. This is simple when tape has been well-placed.
As mentioned above, we recommend the use of a remote shutter or using a short self-timer to minimize vibration-induced blur in the final shots.
|
|
When using OpenCV or Calib.io, consider the following rules of thumb:
- Only 20-30 images are necessary, but more photos will help to remove outliers.
- If possible, estimate only for "k1, k2, k3", leaving out additional p coefficients. High-quality fixed lenses are not expected to have tangential distortion (but this may need to be evaluate for every lens and camera system separately)
- While 30 images may allow you to obtain a reasonable camera calibration, we suggest to take additional photos to allow outlier removal and random selection of photos during advanced processing.
To inform the above recommendations, multiple calibration sessions were conducted with a variety of cameras, targets, image counts, and target-camera orientations. To analyze the success of a calibration, we looked at each calibration's root mean squared reprojection error (RMSE) and distortion plot. To compare two calibrations' distortion plots, we simply took the difference between the two.
To optimize the calibration process, we investigated the performance of estimating for different combinations of distortion coefficients (left), the number of images that should be included in the calibration (center-left), target type (center-right), and camera type (right).
As can be seen in Figure 1, the following potential recommendations emerge:
- Include between 20 and 30 images.
- Use the Calib.io checkerboard (referred to as "chessboard" above), though the Calib.io CharuCo board also performs well.
- While it appears that using "k1, k2, k3", "k1, k2, k3, p1", or "k1, k2, p2" is preferred due to their low RMSE, we decided to look further into the corresponding distortion plots to determine if the additional distortion coefficients were necessary to achieve a good calibration.
We conducted calibrations comparing different combinations of distortion coefficients, and found that there were diminishing returns (and the possibility of overfitting) when including coefficients beyond "k1, k2, k3".
Our results after solving for different combinations of distortion coefficients (using Calib.io). Left column: k1, k2; middle column: other coefficient combinations; right column: comparison plots.
The previous figure shows that, while the combinations of "k1, k2, k3, p1", and "k1, k2, p2 result in slightly lower RMSE than "k1, k2", it is not enough to justify estimating for the additional p coefficients. Therefore, our recommendations are to use simply "k1, k2, k3", without the need for additional p coefficients.
We provide a simple approach to visualize single camera calibration fields:
python visualizeDistortion_from_CC_xml.py
usage: visualizeDistortion_from_CC_xml.py [-h] --png_fn PNG_FN --CC0_fn CC0_FN --title0 TITLE0 --CC1_fn CC1_FN --title1 TITLE1 --CC2_fn CC2_FN --title2 TITLE2
--CC3_fn CC3_FN --title3 TITLE3 [--h H] [--w W]
Visualize up to 4 distortion patterns from camera calibration files ([email protected])
options:
-h, --help show this help message and exit
--png_fn PNG_FN Outputfile for PNG (600 dpi)
--CC0_fn CC0_FN OpenCV CameraCalibration XML file
--title0 TITLE0 Title for plot
--CC1_fn CC1_FN OpenCV CameraCalibration XML file
--title1 TITLE1 Title for plot
--CC2_fn CC2_FN OpenCV CameraCalibration XML file
--title2 TITLE2 Title for plot
--CC3_fn CC3_FN OpenCV CameraCalibration XML file
--title3 TITLE3 Title for plot
--h H Image height (pixels)
--w W Image width (pixels)
Here we show an example of showing the difference between two camera calibration files - one generated from checkerboard and a second one has been optimized within Metashape Agisoft in a setup with a strong camera geometry.
python /home/bodo/Dropbox/soft/github/CameraCalibration/python/compareDistortion_from_CC_xml.py \
--CC0_fn xml/Sony_ILCE-7RM5_50mm_checkerboard_fixed_5p_25Apr2025.xml --title0 'fixed checkerboard 5p' \
--CC1_fn xml/Sony_ILCE-7RM5_50mm_checkerboard_fixed_5p_25Apr2025_adjusted_outside_1_123images_202ktiepoints.xml --title1 'optimized by Metashape Agisoft' \
--save_fname_2panel Sony_ILCE-7RM5_50mm_checkerboard_fixed_MetashapeOptimized_5p_25Apr2025_2panel.png \
--save_fname_2panel_diff Sony_ILCE-7RM5_50mm_checkerboard_fixed_MetashapeOptimized_5p_25Apr2025_2panel_diff.png \
--save_fname_1panel Sony_ILCE-7RM5_50mm_checkerboard_fixed_MetashapeOptimized_5p_25Apr2025_2panel.png \
--title_diff 'Offset Difference fixed checkerboard and Metashape optimized 5 p' \
--suptitle 'Sony 7RM5 50 mm'
Distortion pattern with direction indicated by arrows and distortion magnitude (or offset) shown by color and contour lines.
Difference between the distortion model generated from a fixed checkerboard setting and a Metashape Agisoft optimized camera calibration from a scene with strong camera geometry. Results for a 50 mm lens on a Sony 7RM5 are shown. The optimization process in Agisoft Metashape only slightly changed the camera calibration parameters.