I am trying to upgrade OpenCV libraries on our software from 2.4 to 3.2, but from what I have encountered up until now it seems it's still under heavy development.
With decision trees I straight away got a crash as member cv_labels vector of DTreesImpl class was not initialized, but in calculation it is used as initialized. I used a solution proposed in this comment to proceed further:
(should be a link to github issue, but my karma does not allow to share links :))
void DTreesImpl::startTraining( const Ptr<TrainData>& data, int )
{
...
else
data->getResponses().copyTo(w->ord_responses);
// ---- new code ----
const int n_cv = params.getCVFolds();
if ( n_cv > 0 ) {
int nsamples = (int) w->cat_responses.size();
w->cv_labels.resize( nsamples );
randu( w->cv_labels, 0, n_cv );
}
}
Then I noticed that different decision tree is being generated than what used to be. Tracking it down I noticed that current implementation does not prune trees at all:
int maxdepth = INT_MAX;//pruneCV(root);
Parameter set by setTruncatePrunedTree is not used as well.
Besides I find some node information being hidden from public access in the API but this is minor.
Now I haven't gone too deep further and I haven't tried looking into implementation of other statistical models if there are similar problems, but my question is: Does anybody know what actual state is with machine learning module? Is it still experimental? Is anybody working on it at the moment, because this crash problem exists already 2 years and was not fixed with latest OpenCV release?