[ros-users] pcl model coefficients definitions

Radu Bogdan Rusu rusu at willowgarage.com
Thu Sep 2 07:05:13 UTC 2010


The sac_model_registration replaces sac_registration, and it can be used to find transformations between partially 
overlapping datasets (I just removed sac_registration in trunk which was obsolete - thanks for reminding me). There's a 
tutorial for these things in pcl_tutorials, and I can explain more offline how to tune it for your application if you want.

The non/ linear registration has also been rewritten to use sac_model_registration internally for outlier rejection at 
each iteration. We're still working on "robustifying" it, but it already works well for the datasets that we tried it on.

The reason why sac_model_registration is under sample_consensus is that I wanted to show that with a general purpose 
xSAC (x=RAN,RRAN,MLE,M,RM,etc) framework all you need to do is define your models in terms of:
- sample selection
- distance to model estimation
- inlier estimation (how should a point be considered inlier or not)

and the rest is taken care of for you :)

Cheers,
Radu.

On 09/01/2010 11:49 PM, Adam Leeper wrote:
> I just put a link, though I realized it may have been cleaner to just
> put the stuff I put in the tutorial on the main page instead...  *shrug*
>
> Is sac_registration supposed to be used to find the transformation
> between two identical but transformed point clouds where corresponding
> points have the same index? Or can it be used for something more
> general, like a sensed cloud being matched to a cloud generated from a
> mesh? I'm trying to figure out what it is for/where it could be useful.
>
> Thanks.
>
>
> Adam Leeper
> Stanford University
> aleeper at stanford.edu <mailto:aleeper at stanford.edu>
> 719.358.3804
>
>
> On Wed, Sep 1, 2010 at 11:33 PM, Radu Bogdan Rusu <rusu at willowgarage.com
> <mailto:rusu at willowgarage.com>> wrote:
>
>
>
>     On 09/01/2010 11:23 PM, Adam Leeper wrote:
>
>         I put an edited portion of model_types.h on the wiki, along with my
>         in-line comments about the number and types of model coefficients.
>
>         http://www.ros.org/wiki/pcl/Tutorials/sac_model_coefficients
>
>
>     Thanks a lot! Super useful! Can you link it to the main web page too
>     where we have the models explained? We really need to change the
>     structure of these web pages... i'm starting to dislike the wiki a
>     bit, but we'll find ways around it. Maybe we need to define some
>     sections (per library: Filtering, Feature Estimation, Segmentation,
>     etc) and then write separate pages for each.
>
>
>         After digging around in the source for a while I've come to the
>         conclusion that some of the segmentation models are not
>         implemented. Is
>         that true, or did I miss them somewhere?
>
>
>     You're right. Cone and torus are placeholders right now. It's super
>     hard to find perfect shapes like this in the world, and their
>     mathematical models are more complicated. :( Even with synthetic
>     data, a cylinder model will always respond stronger than a cone
>     model in most cases... We could look into reimplementing them
>     however if there's great interest.
>
>
>         Also, after digging around in the source, it made me appreciate even
>         more how cool pcl is. Nice job Radu :)
>
>
>     Many thanks. :bow: :)
>
>     Cheers,
>     Radu.
>
>     --
>     | Radu Bogdan Rusu | http://rbrusu.com/
>
>

-- 
| Radu Bogdan Rusu | http://rbrusu.com/



More information about the ros-users mailing list