Data Collection for Non Overlapping Camera Calibration
Calibration Target:
Calibration Target:
Checkerboard is the calibration target. Any checkerboard of different sizes and different internal corners can be used. Eg: You can use the following attached pdf. It has 7 internal corners horizontally and 9 internal corners vertically.
Data Collection:
With the support camera and the 2 non overlapping cameras in a fixed position, take images of a checkerboard from the 2 pairs, camera 1 and the support camera and camera 2 and the support camera. Take shots of the checkerboard at different positions with both the pair of cameras. A minimum of 25 images is recommended to get good results. Save the images in different folders on your PC.
Also, record the length of one of the squares on the checkerboard, and also the count of horizontal and vertical corners inside the checkerboard.
It is recommended that you name the files with pairing information in the name itself. For example, name the image from the left and right cameras as ‘left01.jpg’ and ‘right01.jpg’ in case of jpg files, and so on. Keep in mind, there is no need to change the extension while renaming the files. This step is purely to help you pair the images when you upload the images to the calibration tool.
Here is an example set of files used during this calibration process. Here, the count of horizontal corners is 9 and vertical corners are 7.
It is recommended to also have calculated the intrinsics parameters for the cameras to be filled in the Non-overlapping-camera-calibration App with data specifically collected for intrinsic calibration of those cameras. Although an on the fly calculation of intrinsics is supported by the tool, it uses the images uploaded for non-overlapping-camera-calibration, which tends to raise inaccuracies in the intrinsics calibration and overall Non-Overlapping-calibration results.


By definition, this method does not require you to place any target like the checkerboard in the field of view. However, it is recommended that the scene has sufficient objects, textures and unique patterns for feature detectors to identify and match. For example, the above example for calibration target is not an ideal dataset for targetless as it has just a checkerboard in front of a plain white wall.
File naming recommendation is the same as for the above methods
Here is an example set of files used during this calibration process.
Last modified 13d ago
Export as PDF
Copy link