LatentSVM DPM models trained by Pedro Felzenszwalb (Bug #1799)

Added by long long yu almost 13 years ago. Updated over 9 years ago.

Status:Open Start date:2012-04-15
Priority:Normal Due date:
Assignee:- % Done:

0%
Category:objdetect
Target version:-
Affected version: Operating System:
Difficulty: HW Platform:
Pull request:

Description

Hi everyone!

I trained one person detector model usingo the matlab code provided by Pedro Felzenszwalb. I converted the mat format model into xml format model in order to test some pictures using opencv latentsvm detector.

So, after some tests, I found out that the opencv version detect much more false positive than matlab version...

The experiment consist of:

1) testing opencv version + trained_model.xml
2) testing matlab version + trained_model.mat

I attached two images...

well, some possible bug would be:

1) the converted model from *.mat into *.xml doens't work.
2) there is a bug in the opencv code.

person_matlab.png - person detector result matlab (181.3 kB) long long yu, 2012-04-15 09:03 pm

person_opencv.png - person detector result opencv (106.4 kB) long long yu, 2012-04-15 09:03 pm

Associated revisions

Revision 2e79dde3
Added by Roman Donchenko over 11 years ago

Merge pull request #1799 from SpecLad:avformat-53-6

History

#1
Updated by Joel Mckay almost 13 years ago

1) the converted model from *.mat into *.xml doens't work.

Unlikely, as the training data-set ROI import program I wrote creates the same issues for numerous SVM experiments (auto-generates VOC format XML files in C++ with no MatLab run-time needed... related ticket: http://code.opencv.org/issues/1750 ).

2) there is a bug in the opencv code.

OpenCV at one time allowed one to setup the SVM training configuration in ways that were never meant to be compatible (assuming it will run). Additionally, if you exceed 20 classes or around 10000 images the example code may not complete training the SVM.
I have been trying to replicate the heisenbug to exclude my development environment, but have yet to collect meaningful bug data for the OpenCV team (takes some time to run).

Best of luck,
J

#2
Updated by Andrey Kamaev almost 13 years ago

  • Category set to objdetect

#3
Updated by long long yu almost 13 years ago

Thank you Andrey!

I will check the svm implementation of Lsvm ;)

#4
Updated by Kirill Kornyakov almost 13 years ago

  • Tracker changed from Feature to Bug

#5
Updated by Johnson chen almost 10 years ago

long long yu wrote:

Thank you Andrey!

I will check the svm implementation of Lsvm ;)

I have the same question when I use the latentsvmdetect in opencv2.10.
I can not detect the object using OpenCV ,while I can detect it in the Matlab.
What's wrong ?
楼主的问题解决了吗？我现在在Matlab中可以检测目标，但是，在OpenCV中无法检测到，或者漏检率很高，这是什么原因？是不是OpenCV的检测程序需要修改参数？？？

#6
Updated by Maksim Shabunin over 9 years ago

Issue has been transferred to GitHub: https://github.com/Itseez/opencv/issues/4346

Also available in: Atom PDF