def assign_nearest_neighbors(vecs1, vecs2, K=2):
checks = 800
flann_params = {'algorithm': 'kdtree', 'trees': 8}
#pseudo_max_dist_sqrd = (np.sqrt(2) * 512) ** 2
pseudo_max_dist_sqrd = 2 * (512**2)
flann = vt.flann_cache(vecs1, flann_params=flann_params)
try:
fx2_to_fx1, _fx2_to_dist = flann.nn_index(vecs2,
num_neighbors=K,
checks=checks)
except pyflann.FLANNException:
print('vecs1.shape = %r' % (vecs1.shape, ))
print('vecs2.shape = %r' % (vecs2.shape, ))
print('vecs1.dtype = %r' % (vecs1.dtype, ))
print('vecs2.dtype = %r' % (vecs2.dtype, ))
raise
fx2_to_dist = np.divide(_fx2_to_dist, pseudo_max_dist_sqrd)
return fx2_to_fx1, fx2_to_dist
def assign_nearest_neighbors(vecs1, vecs2, K=2):
checks = 800
flann_params = {
'algorithm': 'kdtree',
'trees': 8
}
#pseudo_max_dist_sqrd = (np.sqrt(2) * 512) ** 2
pseudo_max_dist_sqrd = 2 * (512 ** 2)
flann = vt.flann_cache(vecs1, flann_params=flann_params)
try:
fx2_to_fx1, _fx2_to_dist = flann.nn_index(vecs2, num_neighbors=K, checks=checks)
except pyflann.FLANNException:
print('vecs1.shape = %r' % (vecs1.shape,))
print('vecs2.shape = %r' % (vecs2.shape,))
print('vecs1.dtype = %r' % (vecs1.dtype,))
print('vecs2.dtype = %r' % (vecs2.dtype,))
raise
fx2_to_dist = np.divide(_fx2_to_dist, pseudo_max_dist_sqrd)
return fx2_to_fx1, fx2_to_dist
def get_dummy_test_vars1(fname1='easy1.png', fname2='easy2.png'):
import utool as ut
from vtool import image as gtool
from vtool import features as feattool
fpath1 = ut.grab_test_imgpath(fname1)
fpath2 = ut.grab_test_imgpath(fname2)
kpts1, vecs1 = feattool.extract_features(fpath1)
kpts2, vecs2 = feattool.extract_features(fpath2)
chip1 = gtool.imread(fpath1)
chip2 = gtool.imread(fpath2)
#chip1_shape = vt.gtool.open_image_size(fpath1)
#chip2_shape = gtool.open_image_size(fpath2)
#dlen_sqrd2 = chip2_shape[0] ** 2 + chip2_shape[1]
#testtup = (rchip1, rchip2, kpts1, vecs1, kpts2, vecs2, dlen_sqrd2)
import vtool as vt
checks = 800
flann_params = {
'algorithm': 'kdtree',
'trees': 8
}
#pseudo_max_dist_sqrd = (np.sqrt(2) * 512) ** 2
pseudo_max_dist_sqrd = 2 * (512 ** 2)
flann = vt.flann_cache(vecs1, flann_params=flann_params)
import pyflann
try:
fx2_to_fx1, _fx2_to_dist = flann.nn_index(vecs2, num_neighbors=2, checks=checks)
except pyflann.FLANNException:
print('vecs1.shape = %r' % (vecs1.shape,))
print('vecs2.shape = %r' % (vecs2.shape,))
print('vecs1.dtype = %r' % (vecs1.dtype,))
print('vecs2.dtype = %r' % (vecs2.dtype,))
raise
fx2_to_dist = np.divide(_fx2_to_dist, pseudo_max_dist_sqrd)
fx2_to_ratio = np.divide(fx2_to_dist.T[0], fx2_to_dist.T[1])
ratio_thresh = .625
fx2_to_isvalid = fx2_to_ratio < ratio_thresh
fx2_m = np.where(fx2_to_isvalid)[0]
fx1_m = fx2_to_fx1.T[0].take(fx2_m)
#fs_RAT = np.subtract(1.0, fx2_to_ratio.take(fx2_m))
fm_RAT = np.vstack((fx1_m, fx2_m)).T
fm = fm_RAT
return chip1, chip2, kpts1, kpts2, fm
def assign_nearest_neighbors(vecs1, vecs2, K=2):
import vtool as vt
import pyflann
checks = 800
flann_params = {"algorithm": "kdtree", "trees": 8}
# pseudo_max_dist_sqrd = (np.sqrt(2) * 512) ** 2
# pseudo_max_dist_sqrd = 2 * (512 ** 2)
flann = vt.flann_cache(vecs1, flann_params=flann_params)
try:
fx2_to_fx1, fx2_to_dist = matching.normalized_nearest_neighbors(flann, vecs2, K, checks)
# fx2_to_fx1, _fx2_to_dist = flann.nn_index(vecs2, num_neighbors=K, checks=checks)
except pyflann.FLANNException:
print("vecs1.shape = %r" % (vecs1.shape,))
print("vecs2.shape = %r" % (vecs2.shape,))
print("vecs1.dtype = %r" % (vecs1.dtype,))
print("vecs2.dtype = %r" % (vecs2.dtype,))
raise
# fx2_to_dist = np.divide(_fx2_to_dist, pseudo_max_dist_sqrd)
return fx2_to_fx1, fx2_to_dist
def get_dummy_test_vars1(fname1='easy1.png', fname2='easy2.png'):
import utool as ut
from vtool import image as gtool
from vtool import features as feattool
fpath1 = ut.grab_test_imgpath(fname1)
fpath2 = ut.grab_test_imgpath(fname2)
kpts1, vecs1 = feattool.extract_features(fpath1)
kpts2, vecs2 = feattool.extract_features(fpath2)
chip1 = gtool.imread(fpath1)
chip2 = gtool.imread(fpath2)
#chip1_shape = vt.gtool.open_image_size(fpath1)
#chip2_shape = gtool.open_image_size(fpath2)
#dlen_sqrd2 = chip2_shape[0] ** 2 + chip2_shape[1]
#testtup = (rchip1, rchip2, kpts1, vecs1, kpts2, vecs2, dlen_sqrd2)
import vtool as vt
checks = 800
flann_params = {'algorithm': 'kdtree', 'trees': 8}
#pseudo_max_dist_sqrd = (np.sqrt(2) * 512) ** 2
pseudo_max_dist_sqrd = 2 * (512**2)
flann = vt.flann_cache(vecs1, flann_params=flann_params)
import pyflann
try:
fx2_to_fx1, _fx2_to_dist = flann.nn_index(vecs2,
num_neighbors=2,
checks=checks)
except pyflann.FLANNException:
print('vecs1.shape = %r' % (vecs1.shape, ))
print('vecs2.shape = %r' % (vecs2.shape, ))
print('vecs1.dtype = %r' % (vecs1.dtype, ))
print('vecs2.dtype = %r' % (vecs2.dtype, ))
raise
fx2_to_dist = np.divide(_fx2_to_dist, pseudo_max_dist_sqrd)
fx2_to_ratio = np.divide(fx2_to_dist.T[0], fx2_to_dist.T[1])
ratio_thresh = .625
fx2_to_isvalid = fx2_to_ratio < ratio_thresh
fx2_m = np.where(fx2_to_isvalid)[0]
fx1_m = fx2_to_fx1.T[0].take(fx2_m)
#fs_RAT = np.subtract(1.0, fx2_to_ratio.take(fx2_m))
fm_RAT = np.vstack((fx1_m, fx2_m)).T
fm = fm_RAT
return chip1, chip2, kpts1, kpts2, fm
def vsone_feature_matching(kpts1, vecs1, kpts2, vecs2, dlen_sqrd2, cfgdict={},
flann1=None, flann2=None, verbose=None):
r"""
Actual logic for matching
Args:
vecs1 (ndarray[uint8_t, ndim=2]): SIFT descriptors
vecs2 (ndarray[uint8_t, ndim=2]): SIFT descriptors
kpts1 (ndarray[float32_t, ndim=2]): keypoints
kpts2 (ndarray[float32_t, ndim=2]): keypoints
Ignore:
>>> from vtool.matching import * # NOQA
%pylab qt4
import plottool as pt
pt.imshow(rchip1)
pt.draw_kpts2(kpts1)
pt.show_chipmatch2(rchip1, rchip2, kpts1, kpts2, fm=fm, fs=fs)
pt.show_chipmatch2(rchip1, rchip2, kpts1, kpts2, fm=fm, fs=fs)
"""
import vtool as vt
import pyflann
from vtool import spatial_verification as sver
#import vtool as vt
sver_xy_thresh = cfgdict.get('sver_xy_thresh', .01)
ratio_thresh = cfgdict.get('ratio_thresh', .625)
refine_method = cfgdict.get('refine_method', 'homog')
symmetric = cfgdict.get('symmetric', False)
K = cfgdict.get('K', 1)
Knorm = cfgdict.get('Knorm', 1)
#ratio_thresh = .99
# GET NEAREST NEIGHBORS
checks = 800
#pseudo_max_dist_sqrd = (np.sqrt(2) * 512) ** 2
#pseudo_max_dist_sqrd = 2 * (512 ** 2)
if verbose is None:
verbose = True
flann_params = {'algorithm': 'kdtree', 'trees': 8}
if flann1 is None:
flann1 = vt.flann_cache(vecs1, flann_params=flann_params, verbose=verbose)
#print('symmetric = %r' % (symmetric,))
if symmetric:
if flann2 is None:
flann2 = vt.flann_cache(vecs2, flann_params=flann_params, verbose=verbose)
try:
try:
num_neighbors = K + Knorm
fx2_to_fx1, fx2_to_dist = normalized_nearest_neighbors(flann1, vecs2, num_neighbors, checks)
#fx2_to_fx1, _fx2_to_dist = flann1.nn_index(vecs2, num_neighbors=K, checks=checks)
if symmetric:
fx1_to_fx2, fx1_to_dist = normalized_nearest_neighbors(flann2, vecs1, K, checks)
except pyflann.FLANNException:
print('vecs1.shape = %r' % (vecs1.shape,))
print('vecs2.shape = %r' % (vecs2.shape,))
print('vecs1.dtype = %r' % (vecs1.dtype,))
print('vecs2.dtype = %r' % (vecs2.dtype,))
raise
if symmetric:
is_symmetric = flag_symmetric_matches(fx2_to_fx1, fx1_to_fx2)
fx2_to_fx1 = fx2_to_fx1.compress(is_symmetric, axis=0)
fx2_to_dist = fx2_to_dist.compress(is_symmetric, axis=0)
assigntup = assign_unconstrained_matches(fx2_to_fx1, fx2_to_dist)
fx2_match, fx1_match, fx1_norm, match_dist, norm_dist = assigntup
fm_ORIG = np.vstack((fx1_match, fx2_match)).T
fs_ORIG = 1 - np.divide(match_dist, norm_dist)
# APPLY RATIO TEST
fm_RAT, fs_RAT, fm_norm_RAT = ratio_test(fx2_match, fx1_match, fx1_norm,
match_dist, norm_dist,
ratio_thresh)
# SPATIAL VERIFICATION FILTER
#with ut.EmbedOnException():
match_weights = np.ones(len(fm_RAT))
svtup = sver.spatially_verify_kpts(kpts1, kpts2, fm_RAT, sver_xy_thresh,
dlen_sqrd2, match_weights=match_weights,
refine_method=refine_method)
if svtup is not None:
(homog_inliers, homog_errors, H_RAT) = svtup[0:3]
else:
H_RAT = np.eye(3)
homog_inliers = []
fm_RAT_SV = fm_RAT.take(homog_inliers, axis=0)
fs_RAT_SV = fs_RAT.take(homog_inliers, axis=0)
fm_norm_RAT_SV = fm_norm_RAT[homog_inliers]
top_percent = .5
top_idx = ut.take_percentile(fx2_to_dist.T[0].argsort(), top_percent)
fm_TOP = fm_ORIG.take(top_idx, axis=0)
fs_TOP = fx2_to_dist.T[0].take(top_idx)
#match_weights = np.ones(len(fm_TOP))
#match_weights = (np.exp(fs_TOP) / np.sqrt(np.pi * 2))
match_weights = 1 - fs_TOP
#match_weights = np.ones(len(fm_TOP))
svtup = sver.spatially_verify_kpts(kpts1, kpts2, fm_TOP, sver_xy_thresh,
dlen_sqrd2, match_weights=match_weights,
refine_method=refine_method)
if svtup is not None:
(homog_inliers, homog_errors, H_TOP) = svtup[0:3]
np.sqrt(homog_errors[0] / dlen_sqrd2)
else:
H_TOP = np.eye(3)
homog_inliers = []
fm_TOP_SV = fm_TOP.take(homog_inliers, axis=0)
fs_TOP_SV = fs_TOP.take(homog_inliers, axis=0)
matches = {
'ORIG' : MatchTup2(fm_ORIG, fs_ORIG),
'RAT' : MatchTup3(fm_RAT, fs_RAT, fm_norm_RAT),
'RAT+SV' : MatchTup3(fm_RAT_SV, fs_RAT_SV, fm_norm_RAT_SV),
'TOP' : MatchTup2(fm_TOP, fs_TOP),
'TOP+SV' : MatchTup2(fm_TOP_SV, fs_TOP_SV),
}
output_metdata = {
'H_RAT': H_RAT,
'H_TOP': H_TOP,
}
except MatchingError:
fm_ERR = np.empty((0, 2), dtype=np.int32)
fs_ERR = np.empty((0, 1), dtype=np.float32)
H_ERR = np.eye(3)
matches = {
'ORIG' : MatchTup2(fm_ERR, fs_ERR),
'RAT' : MatchTup3(fm_ERR, fs_ERR, fm_ERR),
'RAT+SV' : MatchTup3(fm_ERR, fs_ERR, fm_ERR),
'TOP' : MatchTup2(fm_ERR, fs_ERR),
'TOP+SV' : MatchTup2(fm_ERR, fs_ERR),
}
output_metdata = {
'H_RAT': H_ERR,
'H_TOP': H_ERR,
}
return matches, output_metdata
def vsone_feature_matching(kpts1,
vecs1,
kpts2,
vecs2,
dlen_sqrd2,
cfgdict={},
flann1=None,
flann2=None,
verbose=None):
r"""
Actual logic for matching
Args:
vecs1 (ndarray[uint8_t, ndim=2]): SIFT descriptors
vecs2 (ndarray[uint8_t, ndim=2]): SIFT descriptors
kpts1 (ndarray[float32_t, ndim=2]): keypoints
kpts2 (ndarray[float32_t, ndim=2]): keypoints
Ignore:
>>> from vtool.matching import * # NOQA
%pylab qt4
import plottool as pt
pt.imshow(rchip1)
pt.draw_kpts2(kpts1)
pt.show_chipmatch2(rchip1, rchip2, kpts1, kpts2, fm=fm, fs=fs)
pt.show_chipmatch2(rchip1, rchip2, kpts1, kpts2, fm=fm, fs=fs)
"""
import vtool as vt
import pyflann
from vtool import spatial_verification as sver
#import vtool as vt
sver_xy_thresh = cfgdict.get('sver_xy_thresh', .01)
ratio_thresh = cfgdict.get('ratio_thresh', .625)
refine_method = cfgdict.get('refine_method', 'homog')
symmetric = cfgdict.get('symmetric', False)
K = cfgdict.get('K', 1)
Knorm = cfgdict.get('Knorm', 1)
#ratio_thresh = .99
# GET NEAREST NEIGHBORS
checks = 800
#pseudo_max_dist_sqrd = (np.sqrt(2) * 512) ** 2
#pseudo_max_dist_sqrd = 2 * (512 ** 2)
if verbose is None:
verbose = True
flann_params = {'algorithm': 'kdtree', 'trees': 8}
if flann1 is None:
flann1 = vt.flann_cache(vecs1,
flann_params=flann_params,
verbose=verbose)
#print('symmetric = %r' % (symmetric,))
if symmetric:
if flann2 is None:
flann2 = vt.flann_cache(vecs2,
flann_params=flann_params,
verbose=verbose)
try:
try:
num_neighbors = K + Knorm
fx2_to_fx1, fx2_to_dist = normalized_nearest_neighbors(
flann1, vecs2, num_neighbors, checks)
#fx2_to_fx1, _fx2_to_dist = flann1.nn_index(vecs2, num_neighbors=K, checks=checks)
if symmetric:
fx1_to_fx2, fx1_to_dist = normalized_nearest_neighbors(
flann2, vecs1, K, checks)
except pyflann.FLANNException:
print('vecs1.shape = %r' % (vecs1.shape, ))
print('vecs2.shape = %r' % (vecs2.shape, ))
print('vecs1.dtype = %r' % (vecs1.dtype, ))
print('vecs2.dtype = %r' % (vecs2.dtype, ))
raise
if symmetric:
is_symmetric = flag_symmetric_matches(fx2_to_fx1, fx1_to_fx2)
fx2_to_fx1 = fx2_to_fx1.compress(is_symmetric, axis=0)
fx2_to_dist = fx2_to_dist.compress(is_symmetric, axis=0)
assigntup = assign_unconstrained_matches(fx2_to_fx1, fx2_to_dist)
fx2_match, fx1_match, fx1_norm, match_dist, norm_dist = assigntup
fm_ORIG = np.vstack((fx1_match, fx2_match)).T
fs_ORIG = 1 - np.divide(match_dist, norm_dist)
# APPLY RATIO TEST
fm_RAT, fs_RAT, fm_norm_RAT = ratio_test(fx2_match, fx1_match,
fx1_norm, match_dist,
norm_dist, ratio_thresh)
# SPATIAL VERIFICATION FILTER
#with ut.EmbedOnException():
match_weights = np.ones(len(fm_RAT))
svtup = sver.spatially_verify_kpts(kpts1,
kpts2,
fm_RAT,
sver_xy_thresh,
dlen_sqrd2,
match_weights=match_weights,
refine_method=refine_method)
if svtup is not None:
(homog_inliers, homog_errors, H_RAT) = svtup[0:3]
else:
H_RAT = np.eye(3)
homog_inliers = []
fm_RAT_SV = fm_RAT.take(homog_inliers, axis=0)
fs_RAT_SV = fs_RAT.take(homog_inliers, axis=0)
fm_norm_RAT_SV = fm_norm_RAT[homog_inliers]
top_percent = .5
top_idx = ut.take_percentile(fx2_to_dist.T[0].argsort(), top_percent)
fm_TOP = fm_ORIG.take(top_idx, axis=0)
fs_TOP = fx2_to_dist.T[0].take(top_idx)
#match_weights = np.ones(len(fm_TOP))
#match_weights = (np.exp(fs_TOP) / np.sqrt(np.pi * 2))
match_weights = 1 - fs_TOP
#match_weights = np.ones(len(fm_TOP))
svtup = sver.spatially_verify_kpts(kpts1,
kpts2,
fm_TOP,
sver_xy_thresh,
dlen_sqrd2,
match_weights=match_weights,
refine_method=refine_method)
if svtup is not None:
(homog_inliers, homog_errors, H_TOP) = svtup[0:3]
np.sqrt(homog_errors[0] / dlen_sqrd2)
else:
H_TOP = np.eye(3)
homog_inliers = []
fm_TOP_SV = fm_TOP.take(homog_inliers, axis=0)
fs_TOP_SV = fs_TOP.take(homog_inliers, axis=0)
matches = {
'ORIG': MatchTup2(fm_ORIG, fs_ORIG),
'RAT': MatchTup3(fm_RAT, fs_RAT, fm_norm_RAT),
'RAT+SV': MatchTup3(fm_RAT_SV, fs_RAT_SV, fm_norm_RAT_SV),
'TOP': MatchTup2(fm_TOP, fs_TOP),
'TOP+SV': MatchTup2(fm_TOP_SV, fs_TOP_SV),
}
output_metdata = {
'H_RAT': H_RAT,
'H_TOP': H_TOP,
}
except MatchingError:
fm_ERR = np.empty((0, 2), dtype=np.int32)
fs_ERR = np.empty((0, 1), dtype=np.float32)
H_ERR = np.eye(3)
matches = {
'ORIG': MatchTup2(fm_ERR, fs_ERR),
'RAT': MatchTup3(fm_ERR, fs_ERR, fm_ERR),
'RAT+SV': MatchTup3(fm_ERR, fs_ERR, fm_ERR),
'TOP': MatchTup2(fm_ERR, fs_ERR),
'TOP+SV': MatchTup2(fm_ERR, fs_ERR),
}
output_metdata = {
'H_RAT': H_ERR,
'H_TOP': H_ERR,
}
return matches, output_metdata