bundle adjustment problem for arbitrary data set

108 views
Skip to first unread message

kgeorge2

unread,
Feb 13, 2016, 5:08:05 AM2/13/16
to Ceres Solver
Hi,

I am a ceres newbie. But have enjoyed so far. I would appreciate if somebody can help me out here.


The bundle adjustment problems given at the bal cite  (http://grail.cs.washington.edu/projects/bal/ladybug.html) are pre-initialized, 
There is already initial estimates for the camera parameters K,R, and t and the 3d point positions. 

I am trying to do bundle adjustment on an arbitrary data set. I know the K for my camera, and have the feature point matches all set up.
How do I find the initial  estimates of R and t and the 3d point positions?  

For getting the initial estimates of R and t of the cameras, I have two suggestions none of which I have tried. May be I can get some hint/help here.
1)
One possibility is that, if I am using a camera from my mobile phone to take pictures, I could use the camera motion estimation paraphernalia( accelerometer/gyro etc) to 
estimate the relative R and t for each shot. 

2)Another possibility is that, given pairs of camera-s, I can compute the relative R and t between two cameras by the epipolar-constraint.

What do you suggest? Also, I am not sure, how would I be able to get the initial estimates for 3d positions?

thanks
-koshy



Keir Mierle

unread,
Feb 16, 2016, 2:03:52 PM2/16/16
to ceres-...@googlegroups.com
Hi Koshy,

What you're looking for is information on general structure from motion solving. There are several ways to do it, but one classic way is to find 2 cameras that can see at least 5 points between them, then solve for the relative pose between them (5 points relative pose problem). Then you have the R,t for one camera (assuming other camera is the identity). Then you triangulate the points seen by those two cameras to get their 3D locations. Then you can use camera resection to find the pose of other cameras that can see the just-reconstructed points. After you reconstruct new cameras, then you can find new points seen by those cameras that you can use triangulation to solve. Repeat this until everything is solved; then bundle adjust the lot.


You probably want to pick up a copy of Multiview Geometry.

Thanks,
Keir

--
You received this message because you are subscribed to the Google Groups "Ceres Solver" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ceres-solver...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/1f3d332f-1c1f-4b9b-ae3d-6248ec67dd64%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Chris Sweeney

unread,
Feb 16, 2016, 8:01:44 PM2/16/16
to ceres-...@googlegroups.com

There are also some really great open source solutions for structure from motion that will solve for camera poses, 3d point positions, and do bundle adjustment on everything. Check out the OpenMVG or Theia libraries (both use Ceres to perform BA), among others.


Reply all
Reply to author
Forward
0 new messages