dice

dice copied to clipboard

Hello,

Thank you for providing this nice software. I would like to know if anyone have a successful stereo analysis example for a pronounced/curved surface? I am trying to use DICe on pebbles/rocks but I don't find the right parameters to make the analysis work. My result is always a crushed cloud of dots.

The best I've got in terms of calibrations is an RMS error of 0.3863 and an avg epipolar error of 0.1758. But I don't know if it's good or bad. Do you know any scale/reference for these parameters? I'm using a white background target with black dots.

Best, Jimmy

Edit: I can send you my working directory if you need to check something

Hi Jimmy,

Getting stereo correlation to work well on curved surfaces or any surface that isn't a plane has always been a weakness for DICe. We've done a lot of work to support making this better, but haven't had a chance yet to get his polished up. It's definitely high on our list of objectives. We've implemented most of Jeff Helm's facet-based approach to stereo correlation, but haven't integrated yet into the main correlation routines.

If you have some example pictures, I'd be happy to see if I can get it working in the current version of DICe. Otherwise, it might be a few months before the stereo correlation routines have been improved to handle non-planar surfaces.

Thank you for your response! Here is a link to my working directory: https://we.tl/t-YnfB48lZwa Someone seemed to have quite good results in this post: https://github.com/dicengine/dice/issues/164, I will also send him a message to see if he can tell me what he did.

I have taken a look at your input files and here are a couple observations:

1. The camera angle is quite large (more than 60deg), if it's possible, I would make the angle between the stereo camera pair less.
2. The calibration seems to be a little loose. In correlating features matched from the left to the right image, the epipolar error was large. Do you remember what the epipolar error was for the overall calibration?
3. There's a new version of DICe packaged up https://github.com/dicengine/dice/releases/tag/v3.0-beta.1 that has a better cross correlation initializer. Perhaps you will have better success with this version. It doesn't try to fit the points to a plane like the last version did. You will likely need to fix the camera angle issue first, but after that this may work for you.