I am developing an application that calculates sift of images and matches the set of descritors using vlfeat functions vl_sift and vl_ubcmatch. After getting matches and scores it tells whether the two images are taken at same location or not i.e, if the two images are of Colosseum but with different scale and orientation the application tells that these are of same location.
The question is how should I set a threshold parameter in vl_ubcmatch so that I get best matched features and if I use adaptive threshold then what will be the relationship between matches, scores and number of descriptors.
I have already tried peak and edge threshold in vl_sift but not achieve good results. I am already following need advice for good feature matches