I've mentioned yet in this same forum that my position on auto calibration of two Wii Remote cameras is a quite possible thing. Still, somehow I fail to hit the solution, probably for insufficient accuracy of the cams. (which would eventually mean I have to go for another method)

Let me resume. We put two Wii Remotes on a flat table, taking care of only one thing: that these Wii Remotes be in positions symmetrical to each other, compared to an imaginary mirror plane. Or, if you preferred it as an image, it would quite look like this:



In the main system Oxyz, we have two Wii Remotes placed on each side of y-axis, in (0,D,0) and (0,-D,0). These Wii Remotes are also rotated around z-axis for -α and α, respectively.

Now, how I'd think to get D and α from just one frame. Lets say both Wii Remotes see two Wii Sensor Bar points, placed in positions that we'll call U and V. All data we know are the vectors from the cameras towards these points (which we generated from 2D images that cameras see). Lets call them v', v" (towards point V, as seen from both Wii Remotes) and u', u" (towards point U, as seen from both Wii Remotes). These new vectors are expressed in the corresponding rotated coordinate systems O'x'y'z', and O"x"y"z".

In the final score, we should find D and α, so that the lines defined with these vectors and corresponding coordinate systems intersect themselves in points U and V, proving them to be at the correct distance for the points of Wii Sensor Bar.

First thing to notice is that the lines O'v1' and O"v1" are supposed to be coplanar, which gives us pretty straight picture, since we have four points to fit in a plane. The same goes for O'u1' and O"u1". In this stage, we can set D to an arbitrary value, and focus ourselves to finding the angle alpha. Why can we use an arbitrary D here? Since the coplanarity we rely on has property that the y-axis always belong to our plane, we can freely translate along it, without influencing the result for alpha.

Once we find the best fit alpha for both points, they could differ but very little. There we should take the arithmetic mean to get our final alpha.

Now, let's find D. It suffices to find the current distance between U and V, for ours arbitrary chosen D and found alpha. We will get the distance d. Then we should just scale our arbitrary given D with <Wii Sensor Bar length>/d and that should be it.

Let me repeat. Although everything seems fine on paper, it still escapes me to make it work as an application. I guess it is due to the insufficient quality of the images that cams can get. I will be informing you about the future steps and, of course, thank you in advance for any comment(s) cast.