I ended up implementing vl_ubcmatch in python. Short and sweet :)
def find_matches(template_descriptors, current_img_descriptors, match_thresh):
flann_params = dict(algorithm=1, trees=4)
flann = cv2.flann_Index(current_img_descriptors, flann_params)
idx, dist = flann.knnSearch(template_descriptors, 2, params={})
del flann
matches = np.c_[np.arange(len(idx)), idx[:,0]]
pass_filter = dist[:,0]*match_thresh < dist[:,1]
matches = matches[pass_filter]
return matches
If you want a c++ version, this guy extracted the cpp code from vlfeat's mex code in his blog.
In the comments, he gives a link to his cpp file with the code (change the extension from txt to cpp).
Partially, you can use BFMatcher() for the first definition of vl_ubcmatch (http://www.vlfeat.org/matlab/vl_ubcmatch.html). Although not mentioned in the documentation of BFMatcher, it is also available under python (see also <your-opencv-folder>/samples/python2/find_obj.py):
>>> import cv2
>>> ? cv2.BFMatcher
Type: builtin_function_or_method
String Form:<built-in function BFMatcher>
Docstring: BFMatcher([, normType[, crossCheck]]) -> <BFMatcher object>
The second variant (i.e. the pruning of the results with a threshold) you would have to code yourself, this ratio-test is however pretty easy to accomplish see e.g. http://answers.opencv.org/question/4829/how-to-filter-freakbruteforcematcher-result/ .
In general, however, not each vlfeat-method is that easily adaptable with OpenCV, I wish there would be some work to fuse both great libraries...
Asked: 2013-11-03 04:33:15 -0600
Seen: 2,036 times
Last updated: Nov 05 '13