Featured post
optimization - Retrieving the most significant features gained from SIFT / SURF -
i'm using surf extract features images , match them others. problem images have in excess of 20000 features slows down matching crawl.
is there way can extract n significant features set?
i tried computing mser image , use features within regions. gives me reduction anywhere 5% 40% without affecting matching quality negatively, that's unreliable , still not enough.
additionally size image down, seems affect quality of features severely in cases.
surf offers few parameters (hessian threshold, octaves , layers per octave) couldn't find on how changing these affect feature significance.
after researching , testing have found hessian value each feature rough estimate of it's strength, using top n features sorted hessian not optimal.
achieved better results when doing following until number of features below target of n:
- size image down, if overly large
- only features lie in mser regions considered
- for features lie close each other, feature higher hessian considered
- of n features per image want save, 75% features highest hessian values
- the remaining features taken randomly remainder, weighted distribution of hessian values computed through histogram
now need find suitable n, around 1500 seems enough currently.
- Get link
- X
- Other Apps
Comments
Post a Comment