Sunday, June 5, 2016


Two days ago I wrote the umpteenth Stack Overflow reply about detecting and matching a checkerboard-like calibration target in an image.

The reply mentioned an often-rediscovered algorithm for indexing the detected target corner features into a 2D array isomorphic to the physical checkerboard pattern, so they can be correctly addressed by the later steps of the camera calibration software.

And then I received quite a few private "pings" from people asking for references to the algorithm itself. Yet the beauty of the obvious is that one can't quite remember when and where they saw it first, and point to an original reference. I definitely can't - my own version of that algorithm was developed by myself and a colleague at a now-defunct startup back in '99, I think. I have seen similar techniques in various papers and tech reports over the years, but can't find any in a pinch.

So, for future reference, here is a link to an old report by myself and Roman Waupotitsch describing the whole thing in excruciating detail (algorithms in pseudocode! you have been warned). We even tried to submit the draft to a conference, but it wasn't deemed worthy - applications are for mechanics, I guess, etc.

At the time our implementation (in C++ on Win32) was quite robust and performant - definitely superior to what the early versions of the OpenCV code could do. In particular, it was quite robust to missing / occluded corners, and could easily be generalized to the case of multiple targets in the same image - we had an application doing just that.

Its most glaring weakness was the use of a very simple asymmetry in the pattern for resolving the absolute positions of the corners: we just used two checkerboard columns wider than the rest. Worked for us at the time because the middle section of the board was normally in view in our application, but in every successive re-implementation of the thing I have used the clever "crossratio barcode" pattern by Matsunaga and Kanatani.

1 comment: