def _vote_ransac(self, queryset, dataset, mapping, results, dists):
"""
Applies RANSAC to database features and then votes with them.
Note that normally we RANSAC *after* all the cells are combined,
not on the cells individually. Experimentally this approach
works poorly, but it might be useful as an example.
"""
if self.params['num_neighbors'] > 1:
sorted_counts = self._vote_top_n(queryset, dataset, mapping, results, dists)
else:
sorted_counts = self._vote_matchonce(queryset, dataset, mapping, results, dists)
# filters out outliers from counts until
# filtered_votes(ith image) > votes(jth image) for all j != i
# and returns top 10 filtered results
filtered = {}
bound = -1
num_filt = 0
for siftfile, matches in sorted_counts:
if len(matches) < bound or num_filt > 10:
INFO('stopped after filtering %d' % num_filt)
break
num_filt += 1
F, inliers = corr.find_corr(matches)
bound = max(sum(inliers), bound)
pts = np.ndarray(len(matches), np.object)
pts[0:len(matches)] = matches
if any(inliers):
filtered[siftfile] = list(np.compress(inliers, pts))
rsorted_counts = sorted(filtered.iteritems(), key=lambda x: len(x[1]), reverse=True)
if not rsorted_counts:
INFO('W: ransac rejected everything, not filtering')
return sorted_counts
return rsorted_counts
def _vote_ransac(self, queryset, dataset, mapping, results, dists):
"""
Applies RANSAC to database features and then votes with them.
Note that normally we RANSAC *after* all the cells are combined,
not on the cells individually. Experimentally this approach
works poorly, but it might be useful as an example.
"""
if self.params['num_neighbors'] > 1:
sorted_counts = self._vote_top_n(queryset, dataset, mapping,
results, dists)
else:
sorted_counts = self._vote_matchonce(queryset, dataset, mapping,
results, dists)
# filters out outliers from counts until
# filtered_votes(ith image) > votes(jth image) for all j != i
# and returns top 10 filtered results
filtered = {}
bound = -1
num_filt = 0
for siftfile, matches in sorted_counts:
if len(matches) < bound or num_filt > 10:
INFO('stopped after filtering %d' % num_filt)
break
num_filt += 1
F, inliers = corr.find_corr(matches)
bound = max(sum(inliers), bound)
pts = np.ndarray(len(matches), np.object)
pts[0:len(matches)] = matches
if any(inliers):
filtered[siftfile] = list(np.compress(inliers, pts))
rsorted_counts = sorted(filtered.iteritems(),
key=lambda x: len(x[1]),
reverse=True)
if not rsorted_counts:
INFO('W: ransac rejected everything, not filtering')
return sorted_counts
return rsorted_counts
def rerank_ransac(counts, C, Q):
sorted_counts = sorted(counts.iteritems(), key=lambda x: len(x[1]), reverse=True)
if C.amb_cutoff:
assert C.amb_cutoff > 1
old = sorted_counts
sorted_counts = amb_cutoff(C, Q, sorted_counts)
print "amb cutoff: %d -> %d" % (len(old), len(sorted_counts))
if not sorted_counts:
sorted_counts = old
# tuples of (len, siftfile, matches) ordered [(1..), (2..), (3..)]
reranked = []
num_filt = 0
for siftfile, matches in sorted_counts:
# the bottom g items in 'reranked' are in their final order
g = len(reranked) - bisect.bisect(reranked, (len(matches), None, None))
if g >= C.ranking_min_consistent or num_filt >= C.ranking_max_considered:
if C.verbosity > 0:
INFO('stopped after filtering %d' % num_filt)
break
num_filt += 1
# perform ransac
F, inliers = corr.find_corr(matches)
if any(inliers):
m = np.compress(inliers, matches)
bisect.insort(reranked, (len(m), siftfile, m))
if not reranked:
INFO('W: no db matches passed ransac, not filtering')
return condense2(sorted_counts), False
else:
reranked.reverse()
return condense3(reranked), True
def compute_hom(C, Q, ranked_matches, comb_matches):
match1 = any(check_img(C, Q, ranked_matches[0]))
if not C.compute_hom:
if match1:
return
if not C.log_failures:
return
if C.put_into_dirs:
udir = os.path.join(C.resultsdir, Q.name)
else:
udir = C.resultsdir
if not os.path.exists(udir):
os.makedirs(udir)
i = 0
data = {}
for matchedimg, score, pairs in ranked_matches[:C.max_matches_to_analyze]:
i += 1
if C.stop_on_homTrue and data.get('success'):
break # we are done (found good homography)
clat, clon = getlatlonfromdbimagename(C, matchedimg)
matchimgpath = None
# XXX this sucks, since we don't have a db image abstraction
for ext in ['.jpg', '.JPG', '.png', '.PNG']:
p = os.path.join(C.dbdump, '%s%s' % (matchedimg, ext))
if os.path.exists(p):
matchimgpath = p
assert matchimgpath
match = any(check_img(C, Q, ranked_matches[i-1]))
# matches = db_matches
# rematch for precise fit
db_matches = comb_matches[matchedimg + 'sift.txt']
matchsiftpath = os.path.join(C.dbdump, matchedimg + 'sift.txt')
matches = corr.rematch(C, Q, matchsiftpath)
# matches1 = matches
# rp_matches = corr.rematch(C, Q, matchsiftpath)
# concat db matches
matches.extend(db_matches)
# find homography
rsc_matches, H, inliers = corr.find_corr(matches, hom=True, ransac_pass=True, data=data)
rsc_inliers = np.compress(inliers, rsc_matches).tolist()
u = corr.count_unique_matches(rsc_inliers)
if C.drawtopcorr or (not match and C.log_failures):
# draw picture
matchoutpath = os.path.join(udir, Q.name + ';match' + str(i) + ';gt' + str(match) + ';hom' + str(data.get('success')) + ';uniq=' + str(u) + ';inliers=' + str(float(sum(inliers))/len(matches)) + ';' + matchedimg + '.jpg')
# try:
corr.draw_matches(C, Q, matches, rsc_matches, H, inliers, matchimgpath, matchoutpath, matchsiftpath, C.show_feature_pairs)
# except IOError, e:
# INFO(e)
if C.dump_hom:
if C.put_into_dirs:
identifier = str(i)
else:
identifier = Q.name + ':' + str(i)
H = np.matrix(np.asarray(H))
with open(os.path.join(udir, 'homography%s.txt' % identifier), 'w') as f:
print >> f, H
np.save(os.path.join(udir, 'inliers%s.npy' % identifier), rsc_inliers)
### POSIT ###
if C.do_posit:
posit.do_posit(C, Q, rsc_inliers, matchsiftpath, matchimgpath)
