def _rolling_nanmin_1d(a, w=None):
"""
Compute the rolling min for 1-D while ignoring NaNs.
This essentially replaces:
`np.nanmin(rolling_window(T[..., start:stop], m), axis=T.ndim)`
Parameters
----------
a : ndarray
The input array
w : ndarray, default None
The rolling window size
Returns
-------
output : ndarray
Rolling window nanmin.
"""
if w is None:
w = a.shape[0]
half_window_size = int(math.ceil((w - 1) / 2))
return minimum_filter1d(a, size=w)[half_window_size:half_window_size +
a.shape[0] - w + 1]
def _method_df2(nu, op, alpha=1.0, beta=0.5, nmin_filter=10, log=True):
"""
Cost function: Use normalized first order derivative
Parameters:
-----------
- nu [ndarray]: photon group boundaries
- op [ndarray]: spectra
- alpha [float]: in [0.0, 2.0], default 1.0
alpha < 1: low sensitivity to gradients in the spectra
alpha > 1: high sensitivity to gradients in the spectra
"""
from scipy.ndimage.filters import minimum_filter1d
if log:
err = np.abs(np.gradient(np.gradient(np.log10(op))))
else:
err = np.abs(np.gradient(np.gradient(op)))
err /= err.max() # normalising to 1 so pow gives predictive results
err = err**alpha
if not log:
err /= minimum_filter1d(op, nmin_filter)
err /= err.sum()
err_f = (err * (1 - beta) + beta / len(op))
err_f /= err_f.sum()
return err_f
def get_smoothed_running_minimum(timeseries, tau1=30, tau2=100):
result = minimum_filter1d(uniform_filter1d(timeseries,
tau1,
mode='nearest'),
tau2,
mode='reflect')
return result
def dff(C, sig_baseline=10, win_baseline=300, sig_output=3, method='maximin'):
"""
delta F / F using maximin method from Suite2P
inputs: C - neuropil subtracted fluorescence (neurons x timepoints)
outputs dFF - neurons x timepoints
:param C:
:param sig_baseline:
:param win_baseline:
:param sig_output:
:param method:
:return:
"""
if method == 'maximin': # windowed baseline estimation
flow = filters.gaussian_filter(C, [0, sig_baseline])
flow = filters.minimum_filter1d(flow, win_baseline, axis=1)
flow = filters.maximum_filter1d(flow, win_baseline, axis=1)
else:
flow = None
raise NotImplementedError
C -= flow # substract baseline (dF)
C /= flow # divide by baseline (dF/F)
return filters.gaussian_filter(C, [0, sig_output]) # smooth result
def _method_df2(nu, op, alpha=1.0, beta=0.5, nmin_filter=10, log=True):
"""
Cost function: Use normalized first order derivative
Parameters:
-----------
- nu [ndarray]: photon group boundaries
- op [ndarray]: spectra
- alpha [float]: in [0.0, 2.0], default 1.0
alpha < 1: low sensitivity to gradients in the spectra
alpha > 1: high sensitivity to gradients in the spectra
"""
from scipy.ndimage.filters import minimum_filter1d
if log:
err = np.abs(np.gradient(np.gradient(np.log10(op))))
else:
err = np.abs(np.gradient(np.gradient(op)))
err /= err.max() # normalising to 1 so pow gives predictive results
err = err**alpha
if not log:
err /= minimum_filter1d(op, nmin_filter)
err /= err.sum()
err_f = (err*(1-beta)+beta/len(op))
err_f /= err_f.sum()
return err_f
def validate_signal(t, y, t_ref, y_ref, num=1000, dx=20, dy=0.1):
""" Validate a signal y(t) against a reference signal y_ref(t_ref) by creating a band
around y_ref and finding the values in y outside the band
Parameters:
t time of the signal
y values of the signal
t_ref time of the reference signal
y_ref values of the reference signal
num number of samples for the band
dx horizontal width of the band in samples
dy vertical distance of the band to y_ref
Returns:
t_band time values of the band
y_min lower limit of the band
y_max upper limit of the band
i_out indices of the values in y outside the band
"""
from scipy.ndimage.filters import maximum_filter1d, minimum_filter1d
from scipy.interpolate import interp1d
# re-sample the reference signal into a uniform grid
t_band = np.linspace(start=t_ref[0], stop=t_ref[-1], num=num)
# make t_ref strictly monotonic by adding epsilon to duplicate sample times
for i in range(1, len(t_ref)):
while t_ref[i - 1] >= t_ref[i]:
t_ref[i] = t_ref[i] + 1e-13
interp_method = 'linear' if y.dtype == np.float64 else 'zero'
y_band = interp1d(x=t_ref, y=y_ref, kind=interp_method)(t_band)
y_band_min = np.min(y_band)
y_band_max = np.max(y_band)
# calculate the width of the band
if y_band_min == y_band_max:
w = 0.5 if y_band_min == 0 else np.abs(y_band_min) * dy
else:
w = (y_band_max - y_band_min) * dy
# calculate the lower and upper limits
y_min = minimum_filter1d(input=y_band, size=dx) - w
y_max = maximum_filter1d(input=y_band, size=dx) + w
# find outliers
y_min_i = np.interp(x=t, xp=t_band, fp=y_min)
y_max_i = np.interp(x=t, xp=t_band, fp=y_max)
i_out = np.logical_or(y < y_min_i, y > y_max_i)
# do not count outliers outside the t_ref
i_out = np.logical_and(i_out, t > t_band[0])
i_out = np.logical_and(i_out, t < t_band[-1])
return t_band, y_min, y_max, i_out
def compute_sliding_minmax(array, Window, sig=2):
if sig > 0:
Flow = filters.gaussian_filter1d(array, sig)
else:
Flow = array
Flow = filters.minimum_filter1d(Flow, Window, mode='wrap')
Flow = filters.maximum_filter1d(Flow, Window, mode='wrap')
return Flow
def center_baseline(V, sigma=100, window=500):
''' centers V so the baseline is at 0 '''
Flow = filters.gaussian_filter(V.T, [0.,sigma])
Flow = filters.minimum_filter1d(Flow, window)
Flow = filters.maximum_filter1d(Flow, window)
V_centered = (V.T - Flow).T
#V_centered = (V.T - Flow.mean(axis=1)[:,np.newaxis]).T
return V_centered
def find_reps(y, threshold, open_size, close_size):
"""
From the Y profile of a barbell's path, determine the concentric phase of each rep.
The algorithm is as follows:
1. Compute the gradient (dy/dt) of the Y motion
2. Binarize the gradient signal by a minimum threshold value to eliminate noise.
3. Perform 1D opening by open_size using a minimum then maximum filter in series.
4. Perform 1D closing by close_size using a maximum then minimum filter in series.
The result is a step function that is true for every time point that the concentric (+Y) phase of the rep
is being performed.
Parameters
----------
y : (N) array
Y component of the motion of the barbell path.
threshold : float
Miniumum acceptable value of the gradient (dY/dt) to indicate a rep.
Increasing this can help eliminate noise, but may cause a small delay after a rep begins to when it is
counted, therefore underestimating the time to complete a rep.
open_size : int
Minimum threshold of length of time that it takes to complete a rep (in frames).
Increase this if there are false positive spikes in the rep step signal that are small in width.
close_size : int
Minimum length of time that could be between reps.
Increase this if there are false breaks between reps that should be continuous.
Returns
-------
(N) array
Step signal representing when reps are performed. (1 indicates concentric phase of rep, 0 indicates no rep).
"""
ygrad = np.gradient(y)
rep_signal = np.where(ygrad > threshold, 1, 0)
# Opening to remove spikes
rep_signal = maximum_filter1d(minimum_filter1d(rep_signal, open_size),
open_size)
# Closing to connect movements (as in the step up from the jerk)
rep_signal = minimum_filter1d(maximum_filter1d(rep_signal, close_size),
close_size)
return rep_signal
def running_min(X, tau1, tau2):
###DEBUGGING IMPLEMENTATION###
# return minimum_filter1d(X,tau2,mode = 'nearest')
###PREVIOUS IMPLEMENTATION###
mode = 'nearest'
result = minimum_filter1d(uniform_filter1d(X, tau1, mode=mode),
tau2,
mode='reflect')
return result
def dilate_ranges(self, ranges, angle_increment):
# unvectorized
# i_dilation = int(round(self.angle_dilation / angle_increment))
# for i in range(len(ranges)):
# dilated[i] = np.min(ranges[i-i_dilation:i+i_dilation])
i_dilation = int(round(ANGLE_DILATION / angle_increment)) # one-sided
i_dilation = i_dilation * 2 + 1 # two-sided
dilated = minimum_filter1d(ranges, size=i_dilation)
return dilated
def preprocess(F: np.ndarray,
baseline: str,
win_baseline: float,
sig_baseline: float,
fs: float,
prctile_baseline: float = 0.9) -> np.ndarray:
""" preprocesses fluorescence traces for spike deconvolution
baseline-subtraction with window 'win_baseline'
Parameters
----------------
F : float, 2D array
size [neurons x time], in pipeline uses neuropil-subtracted fluorescence
baseline : str
setting that describes how to compute the baseline of each trace
win_baseline : float
window (in seconds) for max filter
sig_baseline : float
width of Gaussian filter in seconds
fs : float
sampling rate per plane
prctile_baseline : float
percentile of trace to use as baseline if using `constant_prctile` for baseline
Returns
----------------
F : float, 2D array
size [neurons x time], baseline-corrected fluorescence
"""
win = int(win_baseline * fs)
if baseline == 'maximin':
Flow = filters.gaussian_filter(F, [0., sig_baseline])
Flow = filters.minimum_filter1d(Flow, win)
Flow = filters.maximum_filter1d(Flow, win)
elif baseline == 'constant':
Flow = filters.gaussian_filter(F, [0., sig_baseline])
Flow = np.amin(Flow)
elif baseline == 'constant_prctile':
Flow = np.percentile(F, prctile_baseline, axis=1)
Flow = np.expand_dims(Flow, axis=1)
else:
Flow = 0.
F = F - Flow
return F
def validate_signal(t, y, t_ref, y_ref, num=1000, dx=20, dy=0.1):
""" Validate a signal y(t) against a reference signal y_ref(t_ref) by creating a band
around y_ref and finding the values in y outside the band
Parameters:
t time of the signal
y values of the signal
t_ref time of the reference signal
y_ref values of the reference signal
num number of samples for the band
dx horizontal width of the band in samples
dy vertical distance of the band to y_ref
Returns:
t_band time values of the band
y_min lower limit of the band
y_max upper limit of the band
i_out indices of the values in y outside the band
"""
from scipy.ndimage.filters import maximum_filter1d, minimum_filter1d
# re-sample the reference signal into a uniform grid
t_band = np.linspace(start=t_ref[0], stop=t_ref[-1], num=num)
# sort out the duplicate samples before the interpolation
m = np.concatenate(([True], np.diff(t_ref) > 0))
y_band = np.interp(x=t_band, xp=t_ref[m], fp=y_ref[m])
y_band_min = np.min(y_band)
y_band_max = np.max(y_band)
# calculate the width of the band
if y_band_min == y_band_max:
w = 0.5 if y_band_min == 0 else np.abs(y_band_min) * dy
else:
w = (y_band_max - y_band_min) * dy
# calculate the lower and upper limits
y_min = minimum_filter1d(input=y_band, size=dx) - w
y_max = maximum_filter1d(input=y_band, size=dx) + w
# find outliers
y_min_i = np.interp(x=t, xp=t_band, fp=y_min)
y_max_i = np.interp(x=t, xp=t_band, fp=y_max)
i_out = np.logical_or(y < y_min_i, y > y_max_i)
# do not count outliers outside the t_ref
i_out = np.logical_and(i_out, t > t_band[0])
i_out = np.logical_and(i_out, t < t_band[-1])
return t_band, y_min, y_max, i_out
def _compute_alw(y, interval):
a, b = interval
if a == b:
return y
y = shift(y, -a, mode="nearest")
b = min(b, len(y))
# compute offset to left end of window from center
width = int(abs(b - a)) + 1
center = width // 2
return minimum_filter1d(y, b - a + 1, mode="nearest", origin=-center)
def filter1d_same(a: np.ndarray, W: int, max_or_min: str, fillna=np.nan):
out_dtype = np.full(0, fillna).dtype
hW = (W - 1) // 2 # Half window size
if max_or_min == 'max':
out = maximum_filter1d(a, size=W, origin=hW)
else:
out = minimum_filter1d(a, size=W, origin=hW)
if out.dtype is out_dtype:
out[:W - 1] = fillna
else:
out = np.concatenate((np.full(W - 1, fillna), out[W - 1:]))
return out
def sliding_interval_filter(ts, size):
"""USGS HYSEP sliding interval method
The USGS HYSEP sliding interval method as described in `Sloto & Crouse, 1996`_.
The flow series is filter with scipy.ndimage.genericfilter1D using numpy.nanmin function
over a window of size `size`
.. _Slot & Crouse, 1996:
Sloto, Ronald A., and Michele Y. Crouse. “HYSEP: A Computer Program for Streamflow Hydrograph Separation and
Analysis.” USGS Numbered Series. Water-Resources Investigations Report. Geological Survey (U.S.), 1996.
http://pubs.er.usgs.gov/publication/wri964040.
:param size:
:param ts:
:return:
"""
# TODO ckeck the presence of nodata
if (ts.isnull()).any():
blocks, nfeatures = label(~ts.isnull())
block_list = [ts[blocks == i] for i in range(1, nfeatures + 1)]
na_df = ts[blocks == 0]
block_bf = [
pd.Series(data=minimum_filter1d(block, size, mode='reflect'),
index=block.index) for block in block_list
]
baseflow = pd.concat(block_bf + [na_df], axis=0)
baseflow.sort_index(inplace=True)
else:
baseflow = pd.Series(data=minimum_filter1d(ts, size, mode='reflect'),
index=ts.index)
quickflow = ts - baseflow
baseflow.name = 'baseflow'
quickflow.name = 'quickflow'
return baseflow, quickflow
def get_seam(image_energy):
# формируем матрицу
for i in range(1, image_energy.shape[0]):
image_energy[i] += minimum_filter1d(image_energy[i - 1], 3)
# выделяем шов с минимальной энергией:
seam_mask = np.ones_like(image_energy, bool)
j_pos = np.argmin(image_energy[-1])
seam_mask[-1, j_pos] = False
for i in range(image_energy.shape[0] - 2, -1, -1):
j_pos += np.argmin(image_energy[i: i + 1, max(0, j_pos - 1): min(j_pos + 2, image_energy.shape[1])]) - \
(j_pos != 0)
seam_mask[i, j_pos] = False
return seam_mask
def preprocess(F,ops):
sig = ops['sig_baseline']
win = int(ops['win_baseline']*ops['fs'])
if ops['baseline']=='maximin':
Flow = filters.gaussian_filter(F, [0., sig])
Flow = filters.minimum_filter1d(Flow, win)
Flow = filters.maximum_filter1d(Flow, win)
elif ops['baseline']=='constant':
Flow = filters.gaussian_filter(F, [0., sig])
Flow = np.amin(Flow)
elif ops['baseline']=='constant_prctile':
Flow = np.percentile(F, ops['prctile_baseline'], axis=1)
Flow = np.expand_dims(Flow, axis = 1)
else:
Flow = 0.
F = F - Flow
return F
def preprocess(F, ops):
""" preprocesses fluorescence traces for spike deconvolution
baseline-subtraction with window 'win_baseline'
Parameters
----------------
F : float, 2D array
size [neurons x time], in pipeline uses neuropil-subtracted fluorescence
ops : dictionary
'baseline', 'win_baseline', 'sig_baseline', 'fs',
(optional 'prctile_baseline' needed if ops['baseline']=='constant_prctile')
Returns
----------------
F : float, 2D array
size [neurons x time], baseline-corrected fluorescence
"""
sig = ops['sig_baseline']
win = int(ops['win_baseline'] * ops['fs'])
if ops['baseline'] == 'maximin':
Flow = filters.gaussian_filter(F, [0., sig])
Flow = filters.minimum_filter1d(Flow, win)
Flow = filters.maximum_filter1d(Flow, win)
elif ops['baseline'] == 'constant':
Flow = filters.gaussian_filter(F, [0., sig])
Flow = np.amin(Flow)
elif ops['baseline'] == 'constant_prctile':
Flow = np.percentile(F, ops['prctile_baseline'], axis=1)
Flow = np.expand_dims(Flow, axis=1)
else:
Flow = 0.
F = F - Flow
return F
def validate_signal(t, y, t_ref, y_ref, num=1000, dx=20, dy=0.1):
""" Validate a signal y(t) against a reference signal y_ref(t_ref)
t time of the signal
y values of the signal
t_ref time of the reference signal
y_ref values of the reference signal
"""
# re-sample the reference signal into a uniform grid
t_band = np.linspace(start=t_ref[0], stop=t_ref[-1], num=num)
# sort out the duplicate samples before the interpolation
m = np.concatenate(([True], np.diff(t_ref) > 0))
y_band = np.interp(x=t_band, xp=t_ref[m], fp=y_ref[m])
y_band_min = np.min(y_band)
y_band_max = np.max(y_band)
# calculate the width of the band
if y_band_min == y_band_max:
w = 0.5 if y_band_min == 0 else np.abs(y_band_min) * dy
else:
w = (y_band_max - y_band_min) * dy
# calculate the lower and upper limits
y_min = minimum_filter1d(input=y_band, size=dx) - w
y_max = maximum_filter1d(input=y_band, size=dx) + w
# find outliers
y_min_i = np.interp(x=t, xp=t_band, fp=y_min)
y_max_i = np.interp(x=t, xp=t_band, fp=y_max)
i_out = np.logical_or(y < y_min_i, y > y_max_i)
return t_band, y_min, y_max, i_out
def analyze_score(
data_base_dir='/playpen/throat/Endoscope_Study/UNC_HN_Laryngoscopy_004/',
fname='opticalflowscore2.txt',
imgnames=None,
score=None):
print 'Performing boundary detection ... '
if imgnames is None:
assert (score is None)
score = []
imgnames = []
fname = data_base_dir + fname
with open(fname, 'r') as ins:
for line in ins:
pair = line.split()
imgname = pair[0]
imgnames.append(imgname)
score.append(float(pair[1]))
folder = data_base_dir + 'images/'
outputfolder = data_base_dir + 'keyframes/'
if os.path.isdir(outputfolder):
rmtree(outputfolder)
os.mkdir(outputfolder)
keyframes = []
keyframeids = []
keyframesfilename = data_base_dir + 'keyframes.txt'
if os.path.exists(keyframesfilename):
os.remove(keyframesfilename)
keyframesfile = open(keyframesfilename, 'w')
for i in range(len(score)):
#os.system('cp ' + folder + imgnames[i] + ' ' + outputfolder + imgnames[i])
keyframesfile.write('%s\n' % imgnames[i])
keyframes.append(imgnames[i])
keyframeids.append(i)
# boundary detection
boundaries = []
ignore_ends = True
#------------------------------------ boundary detection method 2:
if (ignore_ends):
score[:int(len(score) * 0.1)] = np.ones_like(
score[:int(len(score) * 0.1)])
score[int(len(score) * 0.9):] = np.ones_like(
score[int(len(score) * 0.9):])
threshold = np.percentile(score, 1)
scoreextrema = minimum_filter1d(input=score, size=301)
localextrema = []
localextremaids = []
for i in range(len(score)):
if score[i] == scoreextrema[
i] and score[i] < threshold and score[i] < 0.8:
boundaries.append(imgnames[i])
localextrema.append(score[i])
localextremaids.append(i)
#------------------------------------- boundary detection method 3:
# brutely divide the sequence into 100-frame chunks
# stepsize = 100;
# tempid = stepsize - 1;
# while(tempid < len(score)):
# boundaries.append(imgnames[tempid])
# tempid += stepsize
# ----------------------------------- boundary detection results:
#plt.plot(localextremaids, localextrema, 'ro')
print 'number of boundaries = %d' % len(boundaries)
print 'boundaries :',
print boundaries
#print 'boundaries motion values: ',
#print localextrema
#------------------------------------ split video by boundaries
shots = []
boundaryid = 0
shot = []
for i in range(len(keyframes)):
if boundaryid >= len(boundaries):
shot.append(keyframes[i])
continue
if keyframes[i] < boundaries[boundaryid]:
shot.append(keyframes[i])
else:
shots.append(shot)
shot = []
shot.append(keyframes[i])
boundaryid = boundaryid + 1
shots.append(shot)
#print shots
print 'number of shots = %d' % len(shots)
subid = 0
min_shot_length = 40
print 'minimum shot length threshold = %d' % min_shot_length
for i in range(len(shots)):
shot = shots[i]
print 'shot %d: %d frames' % (i, len(shot)),
if len(shot) > min_shot_length:
subfolder = outputfolder + str(subid) + '/'
os.mkdir(subfolder)
for j in shot:
os.system('cp ' + folder + j + ' ' + subfolder + j)
print ' '
subid = subid + 1
else:
print ' discarded'
h_samples, = plt.plot(sampleids, samples, 'r.')
h_scores, = plt.plot(sampleids, samplescores, 'b+')
h_rate, = plt.plot(rate)
h_low, = plt.plot(2600, rate[2600], 'b^')
h_high, = plt.plot(3770, rate[3770], 'r^')
plt.legend([h_samples, h_scores, h_rate, h_low, h_high], [
'Samples', 'Optical Flow Scores', 'Sample Rate', 'Low Motion Part Example',
'High Motion Part Example'
])
plt.show()
#for i in range(len(imgnames)):
# os.system('cp ' + folder + imgnames[i] + ' ' + outputfolder + imgnames[i])
#exit()
scoreminima = minimum_filter1d(input=score, size=4)
localminima = []
keyframes = []
keyframeids = []
keyframesfilename = data_base_dir + 'keyframes.txt'
if os.path.exists(keyframesfilename):
os.remove(keyframesfilename)
keyframesfile = open(keyframesfilename, 'w')
#for i in range(len(score)):
# if score[i] == scoreminima[i]:
# localminima.append(imgnames[i])
# #os.system('cp ' + folder + imgnames[i] + ' ' + outputfolder + imgnames[i])
# keyframesfile.write('%s\n' % imgnames[i])
# keyframes.append(imgnames[i])
# keyframeids.append(i)
def _compute_bounded_globally(x, a):
z1 = minimum_filter1d(x, a, mode="nearest")
z2 = shift(x, -a, cval=TOP)
z3 = compute_and_binary(z2, z1)
z = compute_and_binary(x, z3)
return z
def main(args):
data_base_dir = args.dir
data_dir = data_base_dir + "images-raw/"
outputfolder = data_base_dir + 'keyframes/'
min_shot_length = args.min_shot_length
nmins_window_size = args.nmins_window_size
percentile_threshold = args.percentile_threshold
absolute_threshold = args.absolute_threshold
ignore_ends = args.ignore_ends
if os.path.isdir(data_base_dir + 'classifiedgood'):
rmtree(data_base_dir + 'classifiedgood')
os.mkdir(data_base_dir + 'classifiedgood')
if os.path.isdir(data_base_dir + 'classifiedbad'):
rmtree(data_base_dir + 'classifiedbad')
os.mkdir(data_base_dir + 'classifiedbad')
if os.path.isdir(outputfolder):
rmtree(outputfolder)
os.mkdir(outputfolder)
#======================
#Load net parames
#======================
style_weights = "./model/weights.pretrained.caffemodel"
test_net = caffe.Net(style_net(train=False, learn_all=False),
style_weights, caffe.TEST)
test_net.forward()
MEAN_FILE = caffe_root + 'data/ilsvrc12/imagenet_mean.binaryproto'
Mean_blob = caffe.proto.caffe_pb2.BlobProto()
Mean_blob.ParseFromString(open(MEAN_FILE, 'rb').read())
# will mean blob to numpy.array
Mean_npy = np.array(caffe.io.blobproto_to_array(Mean_blob))[0]
Mean_npy = Mean_npy.mean(1).mean(
1) # average over pixels to obtain the mean (BGR) pixel values
print 'mean-subtracted values:', zip('BGR', Mean_npy)
# create transformer for the input called 'data'
transformer = caffe.io.Transformer(
{'data': test_net.blobs['data'].data.shape})
transformer.set_transpose(
'data', (2, 0, 1)) # move image channels to outermost dimension
transformer.set_mean(
'data', Mean_npy) # subtract the dataset-mean value in each channel
#transformer.set_raw_scale('data', 255) # rescale from [0, 1] to [0, 255]
transformer.set_channel_swap('data',
(2, 1, 0)) # swap channels from RGB to BGR
atup = ('bad', 'good')
style_labels = list(atup)
classifiedgood = []
opticalflowscore = []
sharpness = []
filelist = os.listdir(data_dir)
filelist.sort()
#filelist = filelist[3000:3100]
count = 1
prev_of = None
# ImageNet classification and optical flow motion estimation
for imfile in filelist:
if imfile.endswith(".jpg"):
print 'Classification and Motion Estimation: %d / %d' % (
count, len(filelist))
count = count + 1
im = Image.open(data_dir + imfile)
im = np.array(im, dtype=np.float32)
transformed_image = transformer.preprocess('data', im)
t = disp_preds(test_net,
transformed_image,
style_labels,
k=2,
name='style')
if t[0] == 1: # only do optical flow on frames classified as good
classifiedgood.append(imfile)
os.system("cp " + data_dir + imfile + " " + data_base_dir +
"classifiedgood")
print 'classified as good : ' + imfile
if len(opticalflowscore) == 0:
opticalflowscore.append(0)
cur_of = rgb2gray(im)
else:
cur_of = rgb2gray(im)
flow = cv2.calcOpticalFlowFarneback(
prev_of, cur_of, 0.5, 3, 15, 3, 5, 1.2, 0)
opticalflowscore.append(
(np.sum(np.absolute(flow[..., 0])) +
np.sum(np.absolute(flow[..., 1]))) / cur_of.size)
#if imfile == 'frame2967.jpg':
# print opticalflowscore[-1]
# temp = Image.open(data_dir + classifiedgood[-2])
# temp = np.array(temp, dtype=np.float32)
# temp = rgb2gray(temp)
# print temp - prev_of
# print imfile
prev_of = cur_of
# find the local minima of the optical flow motion estimation
assert len(opticalflowscore) == len(classifiedgood)
localminima = []
localmin = minimum_filter1d(opticalflowscore, size=3)
for i in range(len(opticalflowscore)):
if localmin[i] == opticalflowscore[i]:
print 'optical flow local minimum: ' + classifiedgood[i]
#os.system('cp '+ folder + filelist[i] + ' ' + outputfolder + filelist[i])
localminima.append(classifiedgood[i])
# calculate sharpness
print 'Calculating sharpness ...'
dl = DoubleList()
D = {}
for f in localminima:
dl.append(f)
D[f] = dl.tail
for i in range(len(localminima)):
img = readImage(data_dir + localminima[i])
sobelx = cv2.Sobel(img, cv2.CV_64F, 1, 0, ksize=5)
sobely = cv2.Sobel(img, cv2.CV_64F, 0, 1, ksize=5)
sharpness.append((np.sum(sobelx**2) + np.sum(sobely**2)) / img.size)
order = np.argsort(sharpness)
# use homography estimation to eliminate the redundant candidates
keyframes = []
score = []
for i in range(len(order)):
print 'Homography estimation: %d / %d ' % (i, len(localminima))
if order[i] == 0 or order[i] == len(localminima) - 1:
keyframes.append(localminima[order[i]])
score.append(0.9)
else:
ID = order[i]
PrevName = D[localminima[ID]].prev.data
NextName = D[localminima[ID]].next.data
Prev = readImage(data_dir + PrevName)
Next = readImage(data_dir + NextName)
H = EstimateHomography(img1=Prev,
img2=Next,
use_builtin_ransac=True)
warpped = cv2.warpPerspective(Prev, H,
(Next.shape[1], Next.shape[0]))
s = correlation_coefficient(warpped, Next)
if s > 0.9:
dl.remove_byaddress(D[localminima[ID]])
print 'Redundant: ' + localminima[ID]
else:
keyframes.append(localminima[order[i]])
score.append(s)
print 'Keyframe: ' + localminima[ID]
#if localminima[ID] == 'frame3423.jpg':
# print PrevName
# print NextName
# print score[ID]
# print '-----------------------'
ordkeyframes = np.argsort(keyframes)
keyframes.sort()
score = np.asarray(score)
score = score[np.asarray(ordkeyframes)]
# key frame selection
# optical flow boundary detection
if ignore_ends:
opticalflowscore[:int(len(opticalflowscore) * 0.1)] = np.zeros_like(
opticalflowscore[:int(len(opticalflowscore) * 0.1)])
opticalflowscore[int(len(opticalflowscore) * 0.9):] = np.zeros_like(
opticalflowscore[int(len(opticalflowscore) * 0.9):])
boundaries = []
threshold = np.percentile(opticalflowscore, percentile_threshold)
opticalextrema = maximum_filter1d(opticalflowscore, nmins_window_size)
for i in range(len(opticalflowscore)):
if opticalflowscore[i] == opticalextrema[i] and opticalflowscore[
i] > threshold and opticalflowscore[i] > absolute_threshold:
boundaries.append(classifiedgood[i])
if len(boundaries) == 0 or boundaries[-1] is not classifiedgood[-1]:
boundaries.append(classifiedgood[-1])
shots = []
boundaryid = 0
shot = []
for i in range(len(keyframes)):
if keyframes[i] <= boundaries[boundaryid]:
shot.append(keyframes[i])
else:
shots.append(shot)
shot = []
shot.append(keyframes[i])
boundaryid = boundaryid + 1
shots.append(shot)
print shots
print len(shots)
final_valid_frames = 0
subid = 0
for i in range(len(shots)):
shot = shots[i]
print 'shot %d: %d frames' % (i, len(shot)),
if 0 and len(shot) > min_shot_length:
subfolder = outputfolder + str(subid) + '/'
os.mkdir(subfolder)
for j in shot:
os.system('cp ' + data_dir + j + ' ' + subfolder + j)
print ' '
final_valid_frames = final_valid_frames + len(shot)
subid = subid + 1
else:
print ' discarded'
# output result to file
for i in keyframes:
os.system('cp ' + data_dir + i + ' ' + outputfolder + i)
resultfile = data_base_dir + 'keyframes.txt'
if os.path.exists(resultfile):
os.remove(resultfile)
thefile = open(resultfile, 'w')
for i in keyframes:
thefile.write("%s\n" % i)
# divide the sequence into shots
#sub_sequences = data_base_dir + 'subsequences.txt'
#if os.path.exists(sub_sequences):
# os.remove(sub_sequences)
#subfile = open(sub_sequences, 'w')
#for i in range(len(score)):
# subfile.write('%s %.4f\n' % (keyframes[i], score[i]))
offilename = data_base_dir + 'opticalflowscore.txt'
if os.path.exists(offilename):
os.remove(offilename)
offile = open(offilename, 'w')
for i in range(len(opticalflowscore)):
offile.write('%s %.4f\n' % (classifiedgood[i], opticalflowscore[i]))
print '%d optical flow local minima' % len(localminima)
print 'Selected %d / %d (%.2f%%) frames as keyframes' % (
final_valid_frames, len(filelist),
100.0 * final_valid_frames / float(len(filelist)))
print 'Split the sequence into %d sub-sequences' % subid
from typing import List
from numpy import array, random, cumsum
from scipy.ndimage.filters import gaussian_filter1d, maximum_filter1d, median_filter, minimum_filter, minimum_filter1d, uniform_filter, gaussian_filter, uniform_filter1d
import pylab
if __name__ == "__main__":
size_of_dataset = 400
t = cumsum(random.randint(0, 10, size_of_dataset))
dataset = cumsum(random.normal(0, 1, size_of_dataset))
pylab.figure(figsize=(8, 8), dpi=80)
pylab.scatter(x=t, y=dataset, c="black", s=2, label="dataset")
pylab.plot(t, uniform_filter1d(dataset, size=20), label="Uniform size=20")
pylab.plot(t,
gaussian_filter1d(dataset, sigma=20),
label="Gaussian $\sigma$=20")
pylab.plot(t, minimum_filter1d(dataset, size=20), label="Minimum size=20")
pylab.plot(t, maximum_filter1d(dataset, size=20), label="Maximum size=20")
pylab.plot(t, median_filter(dataset, size=20), label="Median size=20")
pylab.legend(loc='best')
pylab.xlabel("Time")
pylab.ylabel("Value")
pylab.savefig("filter-application.pdf")
pylab.savefig("filter-application.svg")
pass
if os.path.isdir(outputfolder):
rmtree(outputfolder)
os.mkdir(outputfolder)
prev = rgb2gray(mpimg.imread(folder + filelist[0]))
score = np.empty([len(filelist)])
score[0] = 0
for i in range(1, len(filelist)):
print('image %d / %d' % (i, len(filelist)))
cur = rgb2gray(mpimg.imread(folder + filelist[i]))
flow = cv2.calcOpticalFlowFarneback(prev, cur, 0.5, 3, 15, 3, 5, 1.2, 0)
score[i] = np.sum(np.absolute(flow[..., 0])) + np.sum(
np.absolute(flow[..., 1]))
prev = cur
#plt.figure()
#plt.plot(score)
#plt.show()
localminima = []
localmin = minimum_filter1d(score, size=3)
for i in range(len(score)):
if localmin[i] == score[i]:
print 'keyframe ' + filelist[i]
#os.system('cp '+ folder + filelist[i] + ' ' + outputfolder + filelist[i])
localminima.append(filelist[i])
thefile = open('opticalflowlocalminima.txt', 'w')
for i in localminima:
thefile.write("%s\n" % i)
