Training data for leg_detector?

asked 2014-01-21 06:06:28 -0500

433 ●6 ●10 ●18 http://www.cs.mcgill.c...

Looking at the outputs from the leg_detector package, I'm wondering if it could benefit from a more diverse set of training examples. Notably, the detection confidence falls quickly as people move away from the laser and it is fooled easily by round objects like chair legs.

Does anyone know if the original training data is available anywhere? Maybe also the labelling/annotation method that was used?

Relevant links:

here's an unofficial repo of leg_detector/leg_detector2.
this is David Lu's fork of leg_detector
this is the SRS leg_detector fork

answered 2014-01-21 08:24:47 -0500

David Lu

10932 ●123 ●219 ●250 http://MetroRobots.com/

The training data is not available. The people_experimental repo (which I can't find at the moment) is where the code originally came from. My code (the first two links you posted) were a fork of that.

Caroline Pantofaru was the maintainer there, but when I talked to her, she said that the training data was no longer available.

I do have a collection of bags that I cobbled together that I intended to validate and test the leg detector with, but I did not follow up on that project.

Hope this helps.

edit flag offensive delete link

Comments

Yes, thanks! Do you know anything else about it? It looks like the features used are similar to those in People2D (http://www.informatik.uni-freiburg.de/~spinello/people2D.html), eh? Do you know if there's any high-level description of the tracker/detection matcher used?

Angus ( 2014-01-22 13:02:18 -0500 )edit

Caroline I believe said that this paper (http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4209616&tag=1) was the basis for the leg detector.

David Lu ( 2014-01-22 13:30:16 -0500 )edit

Note: some of the hacks in the code are my own, and an improved version would be greatly appreciated by the community. Or at least myself and other HRI/ROS researchers.

David Lu ( 2014-01-22 13:31:14 -0500 )edit

@David Lu Not sure if this is what you mentioned but I see a copied people_experimental repo.