过度拟合时，OpenCV haar训练无限循环的解决方案是什么？

Question

I run Haar training in openCV 2.4.3.

1. I have 10 negative images and about 30 positive (I know that it is small amount).
2. positive samples are organized in 'vec' file and negative (background ) images are described by text file which stores the path's of the images.
3. With such a small amount of images a perfect solution exists (all of the background images classified as false and the real samples as true).

When this perfect solution found, openCV enters an infinite loop inside icvGetHaarTrainingData() function.

It wants to find background images on which the cascade fails but there are no more such images so this function is stuck in for(;;) loop forever.

icvGetBackgroundImage() function, called in this loop, just keeps loading the same images from hard-disk which are rejected by the cascade.

Does someone know of a patch, or a bug fix I can apply to solve this problem?

One possible fix, I came up with, is to let icvGetBackgroundImage() load each file only once since if it was rejected by the cascade once it would be rejected always. But this is a wrong tweak since this same function is used for generating 'vec' files in cvCreateSamples() and in this scenario we do want it to load the same file few times.

Second possible fix - use negative examples in vec file (like positive examples) instead of info path format. This is a way to overcome infinite loop by not activating the problematic callback, but the result is an openCV error that causes a crash. It happens when negative 'vec' file reaches EOF.

Meanwhile I solved those 2 bugs (infinite loop and crash) by changing directly the source code of openCV and recompiling the binaries. I will submit my changes to openCV committee soon. But I can't believe that I am the only one who bumped into such problem...

Answer 1

Posting here The fixed code. Please note that an original openCV license is applied to it as well. https://github.com/DanielHsH/OpenCV-HaarTraining-BugFix