Training CvGBTrees/CvRTrees with Features?

edit

I'm trying to understand how to use keypoints or other non-HOG-based features for object recognition and I'd really like to try out CvGBTrees or another Random Trees based approach for this.

I have seen a few approaches that have used either hand selected points (i.e. in the points_classifier example included with OpenCV) and also that have used random sampling directly within the image to create a single vector which is then used with a CvRTrees.

Can anyone either confirm that features based detection is plausible or that verify that it isn't at all. Reshaping the descriptors returned from SIFT/SURF to use as training sets for CvRTrees doesn't seem possible, but I'm wondering if there's something that I'm not understanding.

Thanks for any advice or pointers to good resources to understand what feature detection techniques will work well with either CvRTrees or CvGBTrees.