Install opencv libs and cmake:
apt-get install build-essential libopencv-dev cmake
Build:
make
When you have a chessboard pattern with 9 x 7 squares, it has 8 x 6 inner corners. That is what the program will look for. There are no 4-edge corners on the outer edge of the board, so these will not be used.
./opencv-findcorners 8 6 image.png > output.json
This will open a display window to show the detected corners: do check them for correctness! Especially when the image has some visible noise, there is some tendency to misdetect the corner location. When this occurs, fix the output file manually.
The output file is suitable for pasting into a multi-image combined
input file for camcalibplanes.
In case too few points are detected, first try specifying a smaller grid size; so instead of 8 6, specify 7 6 or 8 5. If this works, then either forget about the points that are missing, or add that row or column manually using e.g. hugin.
Also by editing the image, you can try to make detection work in some more cases, but this is usually more trouble than it is worth: using hugin will often be quicker.
Some things you need to know:
The image always gets converted into greyscale, so no need to do external editing in color.
The first detection step converts the image into binary black-and-white,
and the intention is to have all squares still present then. With
CALIB_CB_ADAPTIVE_THRESH active, the black-and-white threshold is not
the same for the entire image, but taken as average over neighboring
pixels. This will allow some gradient in lighting, but not too much.
Some editing possibilities:
Increase image contrast, so that white squares are almost white and black squares are almost black; use "gamma" or "levels" tool in an image editor. When squares are unclean due to reflections or uneven lighting, fix them up -- without touching the edges. (Edges should remain somewhat "blurry" to have better subpixel detection.)
Remove everything except the chessboard. No problem to cut off a little from the board's edge squares -- as long as they remain approximately rectangular with 4 neat corners. Make the outer space approximately as white as the white squares, which works best with the neighbor-average for the adaptive threshold.
For fisheye cameras that have the actual lens edge visible inside the picture, it makes much sense to have some pictures with chessboard very close to the edge. That way, the distortion near the edge can be accurately determined. But squares tend to get very compressed near the edge. It may help somewhat to "fix" the outer squares by stretching them a bit over the edge and painting neat white color around them. As long as the inner corners stay the same, results will be perfectly usable. But detection may still fail, and adding a few pixels manually using hugin is quick and easy too.
Detection will find all "contours" = edges in the black-and-white image after thresholding; and it will keep all square contours.
Then it will use the grayscale image to fine-tune the corners to subpixel precision. Therefore you should not use a black-and-white input image unless really necessary.
In some cases, the program will stubbornly refuse to detect some of the clearly visible inner corners. This may be due to the detection algorithm, and appears generally unfixable. In these cases, doing manual detection using hugin is much faster than trying to convince openCV.