CascadeClassifier results differ significantly on Linux compared to Windows

asked 2015-03-02 03:22:27 -0500

JR hahne gravatar image

Hi there,

I am testing the CascadeClassifier from OpenCV using the pre-trained files

  • haarcascade_eye_tree_eyeglasses.xml
  • haarcascade_frontalface_default.xml

I am using the official OpenCV builds for Windows from SourceForge and the precompiled packages on Ubuntu. I am using the same video input and the same source code for my program.

Anyway, I am getting different results on Linux compared to the results on Windows. The results differs only a few pixels but which is still significant in my case as the results of the following processing slightly differs which pretends to compare the results.

Does anybody know the reason for this or has at least any hints why this is the case? Or should this not be the case?

Thanks a lot in advance, Johannes

edit retag flag offensive close merge delete


In short, a year and half ago they stopped auto merging fixes of the 2.4 branch in the 3.0 branch. This leaded to people fixing stuff in the 2.4 branch without adapting it in the master branch and vice versa. This will be one of the reasons. Another can be that the 3.0 branch supports more multithreading backbones and probably rounding errors can occur there.

StevenPuttemans gravatar imageStevenPuttemans ( 2015-03-02 03:42:23 -0500 )edit