Vector Field Consensus (VFC) is a powerful point pairing algorithm developed in 2014 by Jiayi Ma, Ji Zhao, Jinwen Tian, Alan L. Yuille, and Zhuowen Tu.
VFC is similar to algorithms like RANSAC and PROSAC whereby we perform a series of operations on feature matched keypoints in order to determine which set of keypoints are so-called "inliers" and which are "outliers".
Unlike the two previously mentioned however, VFC is extremely robust and tolerant to very high levels of outliers with only a minimal number of inliers. To quote the abstract from their paper:
Our algorithm starts by creating a set of putative correspondences which can contain a very large number of false correspondences, or outliers, in addition to a limited number of true correspondences (inliers).
I have created these Python3 bindings for their algorithm using their C++ code which I modified very slightly to work with native Python types.
NOTE: The algorithm and this package, is targeted at OpenCV users. Users of other frameworks can leverage this package, but they will have to install OpenCV in order to run this algorithm.
NOTE: As of 3/1/2020 I have only tested the installation on Ubuntu-based GNU/Linux distributions.
apt install libopencv-devpip3 install pyvfc
#!/usr/bin/env python3
import cv2
import pyvfc
im1 = cv2.imread('church1.jpg', 0)
im2 = cv2.imread('church2.jpg', 0)
orb = cv2.ORB_create()
vfc = pyvfc.VFC()
kp1, desc1 = orb.detectAndCompute(im1, None)
kp2, desc2 = orb.detectAndCompute(im2, None)
bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
matches = bf.match(desc1, desc2)
out = cv2.drawMatches(im1, kp1, im2, kp2, matches, None)
cv2.imwrite('initial_match.png', out)
pts1 = [kp1[match.queryIdx].pt for match in matches]
pts2 = [kp2[match.trainIdx].pt for match in matches]
if vfc.setData(pts1, pts2):
vfc.optimize()
match_idx = vfc.obtainCorrectMatch()
good_matches = [matches[idx] for idx in match_idx]
out = cv2.drawMatches(im1, kp1, im2, kp2, good_matches, None)
cv2.imwrite('vfc_results.png', out)
print('VFC ratio %d/%d' % (len(match_idx), len(matches)))