HOGDescriptor: human detection dataset

asked 2015-09-14 03:00:58 -0500

cylix gravatar image

Hi,

I have a little question concerning opencv implementation and maybe someone here knows the answer.

I’m using the HOGDescriptor class for human detection and I’d like to know how was trained the openCV SVM for human body detection: is it with the INRIA dataset (http://pascal.inrialpes.fr/data/human/) or something else?

I’ve searched on the opencv documentation and stackoverflow but I haven’t found what I was looking for.

The problematic behind this question is that I don't get good detection results with the OpenCV dataset and I'd like to know if it could be a good idea to try to train my own SVM with another dataset.

edit retag flag offensive close merge delete

Comments

1

I'm not sure but I think getDefaultPeopleDetector() was trained with INRIA. There's another alternative, getDaimlerPeopleDetector(), that as you would imagine was trained with Daimler dataset (careful if you're going to use it, HOG window has to be 48x96)

LorenaGdL gravatar imageLorenaGdL ( 2015-09-14 04:15:19 -0500 )edit
1

@LorenaGdL as far as I know you are correct for the Daimler one, however the default one works well for upfront and backfront facing positions, but not for side views, so I am not really sure it is made based on INRIA.

That being said, HOG+SVM is so outperformed these days ... for efficient pedestrian detection, move to something like DPM, ICF, ACF, ...

StevenPuttemans gravatar imageStevenPuttemans ( 2015-09-14 06:03:11 -0500 )edit

Thanks for your answers! @StevenPuttemans I suppose that none of the alternatives you mentioned are provided by OpenCV?

cylix gravatar imagecylix ( 2015-09-14 06:13:21 -0500 )edit

There was a GSoC that reimplemented the DPM model, but have not tried it before. There is also work being done on ICF and ACF, however people still report bad performance over the original matlab based code. The goal was to let a GSoC work on that too, but I have not seen any progress on it.

StevenPuttemans gravatar imageStevenPuttemans ( 2015-09-14 06:29:00 -0500 )edit

@StevenPuttemans: I never really thought much about it and always considered the first default detector would be based in INRIA, but what you say is indeed true so who knows what they used... About ICF and ACF, what exactly are the problems right now? I was just about to try them, so knowing issues or bad reports would be helpful

LorenaGdL gravatar imageLorenaGdL ( 2015-09-14 06:59:02 -0500 )edit

I have a PhD collegue working on pedestrian detection who says that the OpenCV ICF implementation works for about 75% of the given samples, whereas the original ICF detector in matlab yields 90%+ detections.

StevenPuttemans gravatar imageStevenPuttemans ( 2015-09-14 07:07:49 -0500 )edit
1

@StevenPuttemans: thanks :) Good to know what to expect

LorenaGdL gravatar imageLorenaGdL ( 2015-09-14 07:42:16 -0500 )edit
1

dpm module (formerly known as latentsvm), inria set is provided there already, for VOCXXX models (including another person set see here )

  • still, much slower than HOG, but due to the parts based idea, covers more poses.
berak gravatar imageberak ( 2015-09-14 11:21:30 -0500 )edit