[ros-users] How to compare two point clouds?

[Christian Mueller]

On 02/18/2011 03:39 PM, Raymond Sheh wrote:
> <snip>
>> Hi all,
>>
>> thanks for the replies.
>> I am working at an object categorization project based on 3D information.
>> I would like to classify objects through their shape properties ( from
>> the object point cloud).
>> First I thought, I can try to fit some primitive shapes like cylinders,
>> spheres, cones, etc.
>> However most of the objects are not symmetric and more complex.
>> Hence I would like to create a kind of codebook of point-cloud-parts
>> from the entire point cloud of the object.
>> Later on, these point-cloud-parts could be compared with a query
>> point-cloud of an object in order to find a similarity.
>> However, to do so a reliable point-cloud similarity measure is required.
>> As far as I know, ICP might be not that good for this purpose, since a
>> good initial guess is required for a reliable alignment. Also the points
>> clouds which are supposed to be compared have to be  similar.
>>
>> Christian
>> _______________________________________________
>> ros-users mailing list
>> ros-users at code.ros.org
>> https://code.ros.org/mailman/listinfo/ros-users
> Perhaps have a look at techniques that find and compare the numbers and
> characteristics of various descriptors (rather than the actual positions
> of individual points). I've been out of this space for a little while so
> I'm not sure what keywords are current ...
> "spin images" spring to mind but of course there are many different
> techniques out there.
>
> - Raymond
> _______________________________________________
> ros-users mailing list
> ros-users at code.ros.org
> https://code.ros.org/mailman/listinfo/ros-users
>

Right!,
I also thought about appearance based approaches which use statistical 
features, like width, height, perimeter, curvatures, volume,.... etc. to 
describe an object.
However the selected features might not be expressive enough to conclude 
that an object is a cup, bowl, or box. That is one disadvantage, but 
there are mostly computationally cheap.

One the other side the first approach (generating a codebook of parts 
from a point-cloud of objects) relies strongly on a reliable point-cloud 
similarity measure and I think this is a quite challenging task compared 
to the generation of statistical features.


thanks,
  Christian