Hi guys!
I'm doing object recognision with bag of words with ORB as descriptor and matcher is:
- in first case BFMatcher with NORM_HAMMING and cross-check enabled
- in second case FlannBasedMatcher with LSH params.
I also create SVM classifiers based on BOW's vocabulary. The interesting thing is that classifiers have better accuracy for setup with LSH than brute force matching. By a lot, like 20%. How it can be possible?
Thanks in advance, Mariusz