Hello,

Has anyone prior experience with adapting tod_training to his/her own training data
without fiducial markers?

I am trying to use the object_recognition stack to recognize textured objects from
2d images. The training data I want to use has no fiducial markers. I have successfully
modified the training pipeline to run on my training images without fiducial markers
by

1) supplying my own masks for the region of interest
2) using already segmented point clouds containing only the points of interests
3) supplying my own pose estimations
4) calling tod_training detector
5) calling tod_training f3d_creator

I am not sure about step 3).

* How influential is the pose information in the recognition process,
   i.e. will I have to put a lots of efforts into the pose estimation or will rough estimates be sufficient?

I would of course also be willing to share my experience gained so far.

Best,
-- Julius

see also http://answers.ros.org/question/289/does-tod_training-in-object_recognition-require