similarity transformation estimation using forward motion images

[imk...@gmail.com]

Hello Dr. Sweeney,

I'm working on large scale reconstruction and we added GPS information to the traditional incremental pipeline. There are two cases where GPS info is used:

1. We transform the camera centers to GPS coordinate system by a similarity transformation;

2. In BA, we add an additional cost function that computes the distance between the camera center and GPS data.

Now the problem is that the similarity transformation estimated from the forward motion images could be incorrect, which causes camera poses to be tilted and creates 'underground' 3D points. As shown below.

Is there any good ways to work around this, any suggestion is appreciated.

thanks,

Kai

[Chris Sweeney]

The issue you are describing is that you are trying to fit a similarity transformation to a trajectory that forms a line. In that case there is a rotational ambiguity (any rotation around the line will align the cameras to the line trajectory!). I would recommend adding a constraint to the rotation by either using measurements of the direction of gravity in your images, or by detecting the ground plane to constrain your rotation.

Chris

--
You received this message because you are subscribed to the Google Groups "Theia Vision Library" group.
To unsubscribe from this group and stop receiving emails from it, send an email to theia-vision-lib...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.