def _postprocess(self, est_times, est_labels):
"""Post processes the estimations from the algorithm, removing empty
segments and making sure the lenghts of the boundaries and labels
match."""
if self.in_labels is not None:
est_labels = np.ones(len(est_times) - 1) * -1
# Remove empty segments if needed
est_times, est_labels = U.remove_empty_segments(est_times, est_labels)
assert len(est_times) - 1 == len(est_labels), "Number of boundaries " \
"(%d) and number of labels(%d) don't match" % (len(est_times),
len(est_labels))
return est_times, est_labels
def _postprocess(self, est_idxs, est_labels):
"""Post processes the estimations from the algorithm, removing empty
segments and making sure the lenghts of the boundaries and labels
match."""
# Remove empty segments if needed
est_idxs, est_labels = U.remove_empty_segments(est_idxs, est_labels)
assert len(est_idxs) - 1 == len(est_labels), "Number of boundaries " \
"(%d) and number of labels(%d) don't match" % (len(est_idxs),
len(est_labels))
# Make sure the indeces are integers
est_idxs = np.asarray(est_idxs, dtype=int)
return est_idxs, est_labels
def _postprocess(self, est_idxs, est_labels):
"""Post processes the estimations from the algorithm, removing empty
segments and making sure the lenghts of the boundaries and labels
match."""
# Remove empty segments if needed
est_idxs, est_labels = U.remove_empty_segments(est_idxs, est_labels)
assert len(est_idxs) - 1 == len(est_labels), "Number of boundaries " \
"(%d) and number of labels(%d) don't match" % (len(est_idxs),
len(est_labels))
# Make sure the indeces are integers
est_idxs = np.asarray(est_idxs, dtype=int)
return est_idxs, est_labels
def processFlat(self):
"""Main process.
Returns
-------
est_idxs : np.array(N)
Estimated indeces for the segment boundaries in frames.
est_labels : np.array(N-1)
Estimated labels for the segments.
"""
# C-NMF params
niter = self.config["niters"] # Iterations for the MF and clustering
# Preprocess to obtain features, times, and input boundary indeces
F = self._preprocess()
# Normalize
F = U.normalize(F, norm_type=self.config["norm_feats"])
if F.shape[0] >= self.config["h"]:
# Median filter
F = median_filter(F, M=self.config["h"])
#plt.imshow(F.T, interpolation="nearest", aspect="auto"); plt.show()
# Find the boundary indices and labels using matrix factorization
est_idxs, est_labels = get_segmentation(
F.T, self.config["rank"], self.config["R"],
self.config["rank_labels"], self.config["R_labels"],
niter=niter, bound_idxs=self.in_bound_idxs, in_labels=None)
# Remove empty segments if needed
est_idxs, est_labels = U.remove_empty_segments(est_idxs, est_labels)
else:
# The track is too short. We will only output the first and last
# time stamps
if self.in_bound_idxs is None:
est_idxs = np.array([0, F.shape[0] - 1])
est_labels = [1]
else:
est_idxs = self.in_bound_idxs
est_labels = [1] * (len(est_idxs) + 1)
# Make sure that the first and last boundaries are included
assert est_idxs[0] == 0 and est_idxs[-1] == F.shape[0] - 1
# Post process estimations
est_idxs, est_labels = self._postprocess(est_idxs, est_labels)
return est_idxs, est_labels
def _postprocess(self, est_idxs, est_labels):
"""Post processes the estimations from the algorithm, removing empty
segments and making sure the lenghts of the boundaries and labels
match."""
# Make sure we are using the previously input bounds, if any
if self.in_bound_idxs is not None:
F = self._preprocess()
est_labels = U.synchronize_labels(self.in_bound_idxs, est_idxs,
est_labels, F.shape[0])
est_idxs = self.in_bound_idxs
# Remove empty segments if needed
est_idxs, est_labels = U.remove_empty_segments(est_idxs, est_labels)
assert len(est_idxs) - 1 == len(est_labels), "Number of boundaries " \
"(%d) and number of labels(%d) don't match" % (len(est_idxs),
len(est_labels))
# Make sure the indeces are integers
est_idxs = np.asarray(est_idxs, dtype=int)
return est_idxs, est_labels