CGAL users discussion list

[Simon Giraudot <>]

Hello,

Le 15/10/2018 à 10:13, williamlai3a a écrit :
> Hi,
>
> I am excited to test out the new classification concept, especially the mesh
> classifier.
> However, I have some doubts on the entire classification process:
> 1. Is training dataset a MUST?
Yes it is. You can use classification without training if you use the 
Sum_of_weighted_features classifier, but in that case the parameters are 
not easy at all to select (especially if use a large number of 
features). I would strongly advise using the 
ETHZ_random_forest_classifier with a training set.
> 2. From my understanding, we only needa label one region per label in the
> dataset (just like the example data given). What if my class has more
> variation in terms of shapes? Shall I give more regions labelled in the
> training dataset?
In general, the larger your training set is, the better the results will 
be. It's especially true if you have a wide variety of local geometric 
features among a label: the training set should be as representative as 
possible.
> 2.1 Would colored point cloud benefit for both training and classification?
Yes, colors can be used and may help classifying better (for example 
with the color channel predefined feature: 
https://doc.cgal.org/latest/Classification/classCGAL_1_1Classification_1_1Feature_1_1Color__channel.html 
).

Note that as the same features should be used for training and for 
classification, you can only use colors for classification if you 
trained your classifier with colors first (and vice versa, a classifier 
trained with colors won't be able to handle point sets without colors).
> 3. Are there any good recommendation software for preparing the .ply
> training data? So I can select and label, then export the property as label
> included in the .ply? (I googled and cannot find any useful one).
I don't really know one.

If you just want to experiment, you can use the classification plugin of 
the Polyhedron_3 demo of CGAL: 
https://github.com/CGAL/cgal/tree/master/Polyhedron/demo/Polyhedron
You can't save labels for a mesh for now (although we should probably 
add such functionality - it only works for point sets so far), but you 
can easily select some facets for training and run classification.

You can see an example on how to use it here: 
https://www.youtube.com/watch?v=xLFm8Aw8vuY
(The video uses a point set, but you can achieve the same thing with 
surface meshes along with the selection plugin.)
> 4. From the given example data, it seems that the points / surfaces are
> spatially disconnected quite sharply. Would the classifier be still working
> well if I have a possion reconstructed surface which is fully connected?
Yes, it should also work. In that case, when creating a training set, it 
will help to provide an accurate training set on the connected regions 
where labels change (otherwise the transitions might be messy).
> [Added]
>
> 5. I am working on similar point cloud like Figure 75.7, however, no example
> codes and data which reproduce Figure 75.7 are given / found. It would be
> great if it is also released:
> - guides on how the training data can be prepared
> - what are the parameters yielding results in Figure 75.7
The figure was done with the demo similarly as what is shown on the 
video, using all existing features on 5 scales (default 
behavior/parameters).

Best regards,

--
Simon Giraudot, PhD
R&D Engineer
GeometryFactory - http://geometryfactory.com/