def sift_match(path1, path2, hop_length, abs_thresh=None, ratio_thresh=None, octave_bins=24, n_octaves=9, fmin=40, timeframe=20, cluster_size=1, sr=22050):
# Extract keypoints
kp1, desc1, img1 = sift_file(path1, hop_length, octave_bins=octave_bins, n_octaves=n_octaves, fmin=fmin, sr=sr)
kp2, desc2, img2 = sift_file(path2, hop_length, octave_bins=octave_bins, n_octaves=n_octaves, fmin=fmin, sr=sr)
# Plot keypoint images
fig = plt.figure()
ax1 = fig.add_subplot(2, 1, 1)
plot_keypoint_image(img1, hop_length, octave_bins, fmin, path1, sr=sr)
ax2 = fig.add_subplot(2, 1, 2)
plot_keypoint_image(img2, hop_length, octave_bins, fmin, path2, sr=sr)
# mng = plt.get_current_fig_manager()
# mng.full_screen_toggle()
# BFMatcher with default params
#bf = cv2.BFMatcher()
#matches = bf.knnMatch(desc1, desc2, k=2)
distances, indices = ann.nearest_neighbors(desc1, desc2, k=2)
# Build match objects
logger.info('Building match objects')
matches = []
for i, distance in enumerate(distances):
matches.append(Match(kp1[i], kp2[indices[i][0]], distance[0], distance[1]))
# Filter nearest neighbors
logger.info('Filtering nearest neighbors down to actual matched samples')
timeframe = int((sr/ hop_length) * timeframe)
matches = filter_matches(matches, abs_thresh, ratio_thresh, timeframe, cluster_size)
# Draw matching lines
plot_matches(ax1, ax2, matches)
plt.show(block=False)
return matches
def sift_match(
path1,
path2,
hop_length,
abs_thresh=None,
ratio_thresh=None,
octave_bins=24,
n_octaves=9,
fmin=40,
timeframe=20,
cluster_size=1,
sr=22050,
):
# Extract keypoints
kp1, desc1, img1 = sift_file(path1,
hop_length,
octave_bins=octave_bins,
n_octaves=n_octaves,
fmin=fmin,
sr=sr)
kp2, desc2, img2 = sift_file(path2,
hop_length,
octave_bins=octave_bins,
n_octaves=n_octaves,
fmin=fmin,
sr=sr)
# Plot keypoint images
fig = plt.figure()
ax1 = fig.add_subplot(2, 1, 1)
plot_keypoint_image(img1, hop_length, octave_bins, fmin, path1, sr=sr)
ax2 = fig.add_subplot(2, 1, 2)
plot_keypoint_image(img2, hop_length, octave_bins, fmin, path2, sr=sr)
# mng = plt.get_current_fig_manager()
# mng.full_screen_toggle()
# BFMatcher with default params
# bf = cv2.BFMatcher()
# matches = bf.knnMatch(desc1, desc2, k=2)
distances, indices = ann.nearest_neighbors(desc1, desc2, k=2)
# Build match objects
logger.info("Building match objects")
matches = []
for i, distance in enumerate(distances):
matches.append(
Match(kp1[i], kp2[indices[i][0]], distance[0], distance[1]))
# Filter nearest neighbors
logger.info("Filtering nearest neighbors down to actual matched samples")
timeframe = int((sr / hop_length) * timeframe)
matches = filter_matches(matches, abs_thresh, ratio_thresh, timeframe,
cluster_size)
# Draw matching lines
plot_matches(ax1, ax2, matches)
plt.show(block=False)
return matches
def match(path1, path2, hop_length, wn=0.4, num_peaks=5, thresh=0.001, plot=True):
target_peaks = num_peaks + 2
fig = plt.figure()
ax1 = fig.add_subplot(2, 1, 1)
S1, peaks1 = spectral_peaks(path1, hop_length, wn=wn, num_peaks=target_peaks, plot=plot)
ax2 = fig.add_subplot(2, 1, 2)
S2, peaks2 = spectral_peaks(path2, hop_length, wn=wn, num_peaks=num_peaks, plot=plot)
mng = plt.get_current_fig_manager()
mng.full_screen_toggle()
logger.info("Finding peak distances...")
peak_dists2 = np.diff(peaks2, axis=0)
num_combos = len(list(itertools.combinations(range(target_peaks), num_peaks)))
peak_dists1 = np.empty([num_peaks - 1, peaks1.shape[1] * (num_combos + 1)], dtype=int)
for x, frame in enumerate(peaks1.T):
start = x * (num_combos + 1)
end = (x + 1) * (num_combos + 1) - 1
peak_dists1[:, start:end] = np.array([np.diff(combo) for combo in itertools.combinations(frame, num_peaks)]).T
logger.info("Finding nearest neighbors")
distances, nearest_neighbors = ann.nearest_neighbors(peak_dists2.T, peak_dists1.T, k=1)
# nearest_neighbors = []
# for frame in peak_dists1:
# score = 0
# nn = None
# for i, yframe in enumerate(peak_dists2):
# count = count_matches(frame, yframe)
# if count > score:
# score = count
# nn = i
# nearest_neighbors.append({'score': score, 'index': nn})
logger.info("Drawing lines between matches")
for x, nn in enumerate(nearest_neighbors[:-1]):
if distances[x] < thresh:
if any(n in nearest_neighbors[x + 1 : x + (num_combos * 3)] for n in [nn + 1, nn + 2, nn + 3]):
con = ConnectionPatch(
xyA=(x / num_combos, 0),
xyB=(nn, S2.shape[0]),
coordsA="data",
coordsB="data",
axesA=ax1,
axesB=ax2,
arrowstyle="<-",
linewidth=1,
zorder=999,
)
ax1.add_artist(con)
ax2.set_zorder(-1)
plt.show(block=False)
return nearest_neighbors, distances
def match(path1, path2, hop_length, wn=0.4, num_peaks=5, thresh=0.001, plot=True):
target_peaks = num_peaks + 2
fig = plt.figure()
ax1 = fig.add_subplot(2, 1, 1)
S1, peaks1 = spectral_peaks(path1, hop_length, wn=wn, num_peaks=target_peaks, plot=plot)
ax2 = fig.add_subplot(2, 1, 2)
S2, peaks2 = spectral_peaks(path2, hop_length, wn=wn, num_peaks=num_peaks, plot=plot)
mng = plt.get_current_fig_manager()
mng.full_screen_toggle()
logger.info('Finding peak distances...')
peak_dists2 = np.diff(peaks2, axis=0)
num_combos = len(list(itertools.combinations(range(target_peaks), num_peaks)))
peak_dists1 = np.empty([num_peaks-1, peaks1.shape[1]*(num_combos+1)], dtype=int)
for x, frame in enumerate(peaks1.T):
start = x*(num_combos+1)
end = (x+1)*(num_combos+1) - 1
peak_dists1[:, start:end] = np.array([np.diff(combo) for combo in itertools.combinations(frame, num_peaks)]).T
logger.info('Finding nearest neighbors')
distances, nearest_neighbors = ann.nearest_neighbors(peak_dists2.T, peak_dists1.T, k=1)
# nearest_neighbors = []
# for frame in peak_dists1:
# score = 0
# nn = None
# for i, yframe in enumerate(peak_dists2):
# count = count_matches(frame, yframe)
# if count > score:
# score = count
# nn = i
# nearest_neighbors.append({'score': score, 'index': nn})
logger.info('Drawing lines between matches')
for x, nn in enumerate(nearest_neighbors[:-1]):
if distances[x] < thresh:
if any(n in nearest_neighbors[x+1:x+(num_combos*3)] for n in [nn+1, nn+2, nn+3]):
con = ConnectionPatch(
xyA=(x/num_combos, 0), xyB=(nn, S2.shape[0]),
coordsA='data', coordsB='data',
axesA=ax1, axesB=ax2,
arrowstyle='<-', linewidth=1,
zorder=999
)
ax1.add_artist(con)
ax2.set_zorder(-1)
plt.show(block=False)
return nearest_neighbors, distances
