easy FLANN training matching and more

classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

easy FLANN training matching and more

berker
Hi all,

I know there are many similar questions but I couldn't find a clear answer.

It seems that there are not an easy and direct training/testing classes/functions for flann based training and matching. There was an app during pcl 1.5 (it has gone now!) and there is the tutorial :

http://pointclouds.org/documentation/tutorials/vfh_recognition.php#vfh-recognition

but it still uses its own functions (which is ok, thank you very much for the work/tutorial).

But my questions are:

1 - Do you plan to implement in-built flann training/testing functions like the one in opencv?
2- How about local features? What would differ in code? (I am using the Lai's RGBD object recognition dataset which has over 600 pcd files for an object instance. So I have to make all of these a class of object)
3 - I couldn't resolve where we add the labels to the classes from the tutorial ???
4 -Do you think converting every feature to "opencv mat" and cont. from that library would be easier? I will try many more classifiers so eventually I guess I have to do this ? Any suggestions on this subject is welcome.

Thanks in advance.
Middle East Technical University
Reply | Threaded
Open this post in threaded view
|

Re: easy FLANN training matching and more

Jochen Sprickerhof
Administrator
Hi berker,

* berker <[hidden email]> [2013-11-26 02:57]:
> There was an app during pcl 1.5 (it has gone now!)

I'm not aware of any deleted apps, maybe it was renamed. Can you tell me
which one you are talking about exactly? Also, you can find 1.5 here:
https://github.com/PointCloudLibrary/pcl/tree/pcl-1.5.0

> 1 - Do you plan to implement in-built flann training/testing functions like
> the one in opencv?

I'm not aware of work going on for this at the moment, contributions
welcome.

> 2- How about local features? What would differ in code? (I am using the
> Lai's RGBD object recognition dataset which has over 600 pcd files for an
> object instance. So I have to make all of these a class of object)

We have a number of local features implemented already, see
https://github.com/PointCloudLibrary/pcl/wiki/Overview-and-Comparison-of-Features

> 3 - I couldn't resolve where we add the labels to the classes from the
> tutorial ???

We don't.

> 4 -Do you think converting every feature to "opencv mat" and cont. from that
> library would be easier? I will try many more classifiers so eventually I
> guess I have to do this ? Any suggestions on this subject is welcome.

Note that PCL has 3D features, whereas OpenCVs are 2D, I don't think
converting them would make any sense.

Cheers Jochen
_______________________________________________
[hidden email] / http://pointclouds.org
http://pointclouds.org/mailman/listinfo/pcl-users
Reply | Threaded
Open this post in threaded view
|

Re: easy FLANN training matching and more

berker
Thank you for your response Jochen,

1- I was talking about the VFH_NN_Classifier:

http://docs.pointclouds.org/1.5.1/classpcl_1_1_v_f_h_classifier_n_n.html

http://svn.pointclouds.org/pcl/branches/pcl-1.5.x/apps/src/nn_classification_example.cpp

2 - I know all about the available local features. My question was: according to tutorial would using local features affect the pipeline and code?

3 -Without labels how can you train several classes, maybe I couldn't understand the purpose of the tutorial. I should dig more into the tutorial and flann.

4- I am not planning to convert the point clouds to cv::mat. I am planning to pass the extracted feature vectors / descriptors (which you can consider them to be 1D) to opencv as cv::mat. So that I will be able to use opencv's classifiers on these vectors.
Middle East Technical University