def match_frames(f1, f2):
# Instantiate BFMatcher with OpenCV
bf = cv2.BFMatcher(cv2.NORM_HAMMING)
# Use BFMatcher with k-nearest neighbors matching to find matched points within two images
matches = bf.knnMatch(f1.des, f2.des, k=2)
# Lowe's ratio test
ret = []
idx1, idx2 = [], []
idx1s, idx2s = set(), set()
# Iterate over matches
for m, n in matches:
# Only use points within a certain range
if m.distance < 0.75 * n.distance:
p1 = f1.kps[m.queryIdx]
p2 = f2.kps[m.trainIdx]
# Only use points within ORB distance equal to 32
if m.distance < 32:
# Append points
if m.queryIdx not in idx1s and m.trainIdx not in idx2s:
idx1.append(m.queryIdx)
idx2.append(m.trainIdx)
idx1s.add(m.queryIdx)
idx2s.add(m.trainIdx)
ret.append((p1, p2))
# Assertions
assert (len(set(idx1)) == len(idx1))
assert (len(set(idx2)) == len(idx2))
assert len(ret) >= 8
# Convert lists to numpy arrays
ret = np.array(ret)
idx1 = np.array(idx1)
idx2 = np.array(idx2)
# Identify transformation matrix between to frames using RANSAC algorithm
model, inliers = ransac((ret[:, 0], ret[:, 1]),
EssentialMatrixTransform,
min_samples=8,
residual_threshold=RANSAC_RESIDUAL_THRES,
max_trials=RANSAC_MAX_TRIALS)
# Print model inliers and results
print("Matches: %d -> %d -> %d -> %d" %
(len(f1.des), len(matches), len(inliers), sum(inliers)))
# Return model inliers and model parameters
print("Model params", model.params)
print("...fundamental to R-t", fundamentalToRt(model.params))
return idx1[inliers], idx2[inliers], fundamentalToRt(model.params)
def match_frames(frame_1: Frame, frame_2: Frame):
bf = cv2.BFMatcher(cv2.NORM_HAMMING)
matches = bf.knnMatch(frame_1.descriptors, frame_2.descriptors, k=2)
# Lowe's ratio test
ret = []
idx1, idx2 = [], []
idx1s, idx2s = set(), set()
for m,n in matches:
if m.distance < 0.75*n.distance:
p1 = frame_1.kps[m.queryIdx]
p2 = frame_2.kps[m.trainIdx]
# be within orb distance 32
if m.distance < 32:
# keep around indices
# TODO: refactor this to not be O(N^2)
if m.queryIdx not in idx1s and m.trainIdx not in idx2s:
idx1.append(m.queryIdx)
idx2.append(m.trainIdx)
idx1s.add(m.queryIdx)
idx2s.add(m.trainIdx)
ret.append((p1, p2))
# no duplicates
assert(len(set(idx1)) == len(idx1))
assert(len(set(idx2)) == len(idx2))
if len(ret) < 8:
logger.warning("Skipping match of frame {} to frame {}".format(frame_1.id, frame_2.id))
raise NoFrameMatchError
ret = np.array(ret)
idx1 = np.array(idx1)
idx2 = np.array(idx2)
# fit matrix
model, inliers = ransac((ret[:, 0], ret[:, 1]),
EssentialMatrixTransform,
min_samples=8,
residual_threshold=RANSAC_RESIDUAL_THRES,
max_trials=RANSAC_MAX_TRIALS)
logger.info("Quality: {}".format(np.mean(ret[:, 0]-ret[:, 1])))
logger.info("Matches: %d -> %d -> %d -> %d" % (len(frame_1.descriptors), len(matches), len(inliers), sum(inliers)))
return idx1[inliers], idx2[inliers], fundamentalToRt(model.params)
def match_frames(f1, f2):
bf = cv2.BFMatcher(cv2.NORM_HAMMING)
matches = bf.knnMatch(f1.des, f2.des, k=2)
# Lowe's ratio test
ret = []
idx1, idx2 = [], []
idx1s, idx2s = set(), set()
for m, n in matches:
if m.distance < 0.75 * n.distance:
p1 = f1.kps[m.queryIdx]
p2 = f2.kps[m.trainIdx]
# be within orb distance 32
if m.distance < 32:
# keep around indices
# TODO: refactor this to not be O(N^2)
if m.queryIdx not in idx1s and m.trainIdx not in idx2s:
idx1.append(m.queryIdx)
idx2.append(m.trainIdx)
idx1s.add(m.queryIdx)
idx2s.add(m.trainIdx)
ret.append((p1, p2))
# no duplicates
assert (len(set(idx1)) == len(idx1))
assert (len(set(idx2)) == len(idx2))
assert len(ret) >= 8
ret = np.array(ret)
idx1 = np.array(idx1)
idx2 = np.array(idx2)
# fit matrix
model, inliers = ransac((ret[:, 0], ret[:, 1]),
FundamentalMatrixTransform,
min_samples=8,
residual_threshold=0.001,
max_trials=100)
print("Matches: %d -> %d -> %d -> %d" %
(len(f1.des), len(matches), len(inliers), sum(inliers)))
return idx1[inliers], idx2[inliers], fundamentalToRt(model.params)