Eric Perko wrote:
> Have you tried different baselines to see if that affects the
> calibration results at all? And I'm pretty sure my checkerboard has
> 28.58 mm squares, so yours is not that much smaller. Can you try a
> larger checkerboard?
>
I'd have to print one. I've tried this multiple times and at a variety
of distances.
> I'd assume that, if your cameras are not moving and you can hold the
> checkerboard still until both cameras have taken a pair, syncing
> shouldn't be a huge problem for calibration. What happens if you
> calibrate each camera separately and view rectified images? Do they
> come out okay or are there issues with monocular calibration as well?
> Specifically, does that checkerboard work for monocular calibration?
>
Yes. Monocular works fine.
> Also, some other students working on stereovision using OpenCV
> directly have found that you have to be careful while watching the
> calibration to make sure that the corner detection is working
> properly. For example, if it begins by marking corners in red on top
> in both cameras, it has to always mark the top corners in red - if
> they ever flip or otherwise change orientation, it will make things go
> quite funky. I've seen it happen with my monocular calibrations if the
> person with the checkerboard isn't careful and accidentally lets the
> checkerboard rotate along the axis from the camera to the
> checkerboard.
>
Didn't happen.
Problem is that R and T are retuning VERY high values. I just went from
trunk to latest, added a small fix to the node source to only accept two
images if there timestamps are within 0.05 seconds of each other, and
tried again. With latest, I get even larger crazy values in R. Again,
I can send the output tar file, which I assume you can run through in an
offline manner, and you can see that the checkerboard is proerly found
on all used frames.
(text/plain)