I try to train a Descriptor Database for Natural Feature Detection. To find the features I use FAST Corner Detector and SIFT Descriptor. Because it is possible to that I detect good and bad Features I want to filter them so that I only get the best for the Descriptor calculation, because it takes really long.
A approach I already implemented is, to filter that keypoints, that have more than 3 Peaks in their 15x15 Patch Orientation Histogram. It was easy to implement that but I still get a big Number of Features ~400. AFAIK It would be more than enough if I have <100 Features for the Matching later on. So how can I filter the Features and have only the good ones left for the descriptor calculation? As I see the FAST algorithm also sets the response variable from the KeyPoint object. But is this value meaningful? I thought about sth. like the low contrast filter or edge response filter from Lowe's paper but because I dont use DoG for Detection I cant access that or not?