def match_adalam_with_magsac(kps1, kps2, descs1, descs2, h1, w1, h2, w2):
matcher = AdalamFilter()
kp1, s1, a1, r1 = convert_kpts_to_imc(kps1)
kp2, s2, a2, r2 = convert_kpts_to_imc(kps2)
idxs = matcher.match_and_filter(kp1,
kp2,
descs1,
descs2,
im1shape=(h1, w1),
im2shape=(h2, w2),
o1=a1.reshape(-1),
o2=a2.reshape(-1),
s1=s1.reshape(-1),
s2=s2.reshape(-1)).detach().cpu().numpy()
if len(idxs) <= 15:
return None
src_pts = kp1[idxs[:, 0]]
dst_pts = kp2[idxs[:, 1]]
F, inliers_mask = cv2.findFundamentalMat(src_pts,
dst_pts,
method=cv2.USAC_MAGSAC,
ransacReprojThreshold=0.25,
confidence=0.99999,
maxIters=100000)
if np.array(inliers_mask).sum() < 15:
return None
inliers_mask = np.array(inliers_mask).astype(bool).reshape(-1)
return np.concatenate([src_pts[inliers_mask], dst_pts[inliers_mask]],
axis=1)
def sequential_adalam_matcher(kp1, kp2, desc1, desc2, meta_desc1,
meta_desc2, img_shape1, img_shape2):
device = desc1.device
desc_sims = 0.
weights_sum = 0.
for i in range(4):
weights = torch.exp(meta_desc1[:, i, :] @ meta_desc2[:, i, :].t())
weights_sum += weights
desc_sims += (desc1[:, i, :] @ desc2[:, i, :].t()) * weights
del meta_desc1, meta_desc2, desc1, desc2, weights
desc_sims /= weights_sum
del weights_sum
desc_sims = 2 - 2 * desc_sims
# Compute putative matches and mutual neighbors
dd12, nn12 = torch.topk(desc_sims, k=2, dim=1, largest=False)
putative_matches = nn12[:, 0]
scores = dd12[:, 0] / dd12[:, 1].clamp_min_(1e-3)
dd21, nn21 = torch.min(desc_sims, dim=0)
mnn = nn21[putative_matches] == torch.arange(kp1.shape[0], device=device)
del desc_sims
# Filter the matches with Adalam
matcher = AdalamFilter()
matches = matcher.filter_matches(
kp1, kp2, putative_matches, scores, mnn, img_shape1, img_shape2,
None, None, None, None)
return matches.data.cpu().numpy()
def calculateImages(self, im1, im2):
k1, o1, s1, d1, im1 = self.extract_keypoints(im1)
k2, o2, s2, d2, im2 = self.extract_keypoints(im2)
print("After extracting points, k1 size: ", len(k1))
print("After extracting points, k2 size: ", len(k2))
matcher = AdalamFilter()
matches = matcher.match_and_filter(
k1=k1,
k2=k2,
o1=o1,
o2=o2,
d1=d1,
d2=d2,
s1=s1,
s2=s2,
im1shape=im1.shape[:2],
im2shape=im2.shape[:2]).cpu().numpy()
k1 = k1[matches[:, 0]]
k2 = k2[matches[:, 1]]
F, mask = cv.findFundamentalMat(k1, k2, cv.FM_LMEDS)
print("\n")
print("Fundamental Matrix is: ")
print(F)
print("\n")
pts1 = k1[mask.ravel() == 1]
pts2 = k2[mask.ravel() == 1]
# print("size of pts1 is: ", len(pts1))
file1 = open("Adalam_matched_Left.txt", "w")
file2 = open("Adalam_matched_Right.txt", "w")
for (x1, y1), (x2, y2) in zip(k1, k2):
# print(x1, " ", y1, " ", x2, " ", y2)
file1.write(str(x1) + ", " + str(y1) + "\n")
file2.write(str(x2) + ", " + str(y2) + "\n")
file1.close()
file2.close()
out1 = k1[mask.ravel() == 0]
out2 = k2[mask.ravel() == 0]
self.show_matches(im1, im2, pts1, pts2, out1, out2)
def adalam_matcher(kp1, kp2, desc1, desc2, meta_desc1,
meta_desc2, img_shape1, img_shape2):
device = desc1.device
desc_weights = torch.einsum('nid,mid->nim', (meta_desc1, meta_desc2))
del meta_desc1, meta_desc2
desc_weights = func.softmax(desc_weights, dim=1)
desc_sims = torch.einsum('nid,mid->nim', (desc1, desc2)) * desc_weights
del desc1, desc2, desc_weights
desc_sims = 2 - 2 * torch.sum(desc_sims, dim=1)
# Compute putative matches and mutual neighbors
dd12, nn12 = torch.topk(desc_sims, k=2, dim=1, largest=False)
putative_matches = nn12[:, 0]
scores = dd12[:, 0] / dd12[:, 1].clamp_min_(1e-3)
dd21, nn21 = torch.min(desc_sims, dim=0)
mnn = nn21[putative_matches] == torch.arange(kp1.shape[0], device=device)
del desc_sims
# Filter the matches with Adalam
matcher = AdalamFilter()
matches = matcher.filter_matches(
kp1, kp2, putative_matches, scores, mnn, img_shape1, img_shape2,
None, None, None, None)
return matches.data.cpu().numpy()
import argparse
from adalam import AdalamFilter
if __name__ == '__main__':
p = argparse.ArgumentParser(description="Match a colmap database with AdaLAM")
p.add_argument("--database_path", "-d", required=True)
p.add_argument("--image_pairs_path", "-i", required=True)
opt = p.parse_args()
matcher = AdalamFilter()
matcher.match_colmap_database(database_path=opt.database_path, image_pairs_path=opt.image_pairs_path)
cv.line(vis, (x1, y1), (x2, y2), [0, 255, 0], 1)
cv.imshow("AdaLAM example", vis)
cv.waitKey()
if __name__ == '__main__':
p = argparse.ArgumentParser()
p.add_argument("--im1", required=True)
p.add_argument("--im2", required=True)
opt = p.parse_args()
# results = function(im1, im2)
k1, o1, s1, d1, im1 = extract_keypoints(opt.im1)
k2, o2, s2, d2, im2 = extract_keypoints(opt.im2)
matcher = AdalamFilter()
matches = matcher.match_and_filter(k1=k1,
k2=k2,
o1=o1,
o2=o2,
d1=d1,
d2=d2,
s1=s1,
s2=s2,
im1shape=im1.shape[:2],
im2shape=im2.shape[:2]).cpu().numpy()
show_matches(im1, im2, k1=k1[matches[:, 0]], k2=k2[matches[:, 1]])
# Save as image and rename