def noiseFraction(truth_h5, measured_h5, tolerance):
"""
Return the fraction of measured localizations that are greater than
tolerance pixels from the nearest truth localization.
Note: This will return 0 if there are no measured localizations.
truth_h5 - A saH5Py.SAH5Py object with the ground truth localizations.
measured_h5 - A saH5Py.SAH5Py object with the found localizations.
tolerance - The search radius in pixels.
"""
if (measured_h5.getNLocalizations() == 0):
return [0, truth_h5.getNLocalizations()]
noise_locs = 0
total_locs = 0
for i in range(truth_h5.getMovieLength()):
t_locs = truth_h5.getLocalizationsInFrame(i)
m_locs = measured_h5.getLocalizationsInFrame(i)
if bool(t_locs) and bool(m_locs):
dist = iaUtilsC.peakToPeakDistAndIndex(m_locs['x'],
m_locs['y'],
t_locs['x'],
t_locs['y'],
max_distance=tolerance)[0]
noise_locs += numpy.count_nonzero((dist < 0.0))
total_locs += dist.size
elif bool(m_locs):
noise_locs += m_locs['x'].size
total_locs += m_locs['x'].size
return [noise_locs, total_locs]
def xyDrift(locs_filename):
# Load localizations.
#
with saH5Py.SAH5Py(locs_filename) as h5:
n_frames = h5.getMovieLength()
f0_locs = h5.getLocalizationsInFrame(0, fields = ["x", "y"])
fn_locs = h5.getLocalizationsInFrame(n_frames - 1, fields = ["x", "y"])
assert (f0_locs["x"].size > 0), "No localizations in the first frame."
assert (fn_locs["y"].size > 0), "No localizations in the last frame."
#
# Identify matching beads in the first and last frame and
# compute the displacement.
#
p_index = utilC.peakToPeakDistAndIndex(f0_locs['x'], f0_locs['y'],
fn_locs['x'], fn_locs['y'])[1]
all_dx = []
all_dy = []
for i in range(f0_locs['x'].size):
dx = fn_locs['x'][p_index[i]] - f0_locs['x'][i]
dy = fn_locs['y'][p_index[i]] - f0_locs['y'][i]
if ((dx*dx + dy*dy) < 1.0):
all_dx.append(dx)
all_dy.append(dy)
#
# Return average per frame.
#
return [numpy.mean(numpy.array(all_dx))/float(n_frames),
numpy.mean(numpy.array(all_dy))/float(n_frames)]
def recallFraction(truth_h5, measured_h5, tolerance):
"""
Return the fraction of truth localizations that have a
measured localization within X pixels (in the XY plane).
truth_h5 - A saH5Py.SAH5Py object with the ground truth localizations.
measured_h5 - A saH5Py.SAH5Py object with the found localizations.
tolerance - The search radius in pixels.
"""
if (measured_h5.getNLocalizations() == 0):
return [0, truth_h5.getNLocalizations()]
recalled_locs = 0
total_locs = 0
for i in range(truth_h5.getMovieLength()):
t_locs = truth_h5.getLocalizationsInFrame(i)
m_locs = measured_h5.getLocalizationsInFrame(i)
if bool(t_locs) and bool(m_locs):
dist = iaUtilsC.peakToPeakDistAndIndex(t_locs['x'],
t_locs['y'],
m_locs['x'],
m_locs['y'],
max_distance=tolerance)[0]
recalled_locs += numpy.count_nonzero((dist >= 0.0))
total_locs += dist.size
elif bool(t_locs):
total_locs += t_locs['x'].size
return [recalled_locs, total_locs]
def collateRQE(dirs, settings):
"""
Results collation for RQE correction.
"""
for a_dir in dirs:
print("Processing", a_dir)
t_locs = saH5Py.loadLocalizations("grid_list.hdf5", fields=["x", "y"])
t_locs_found = numpy.zeros_like(t_locs["x"])
n_frames = 0
with saH5Py.SAH5Py(os.path.join(a_dir, "test.hdf5")) as h5:
for i in range(h5.getMovieLength()):
n_frames += 1
m_locs = h5.getLocalizationsInFrame(i, fields=["x", "y"])
dist = iaUtilsC.peakToPeakDistAndIndex(t_locs['x'],
t_locs['y'],
m_locs['x'],
m_locs['y'],
max_distance=3)[0]
for j in range(dist.size):
if (dist[j] > -0.1):
t_locs_found[j] += 1
# Check results against the binomial distribution.
p = numpy.sum(t_locs_found) / (n_frames * t_locs["x"].size)
print(" Mean P found : {0:.3f}".format(p))
print(" Expected variance: {0:.3f}".format(n_frames * p * (1 - p)))
print(" Actual variance : {0:.3f}".format(numpy.var(t_locs_found)))
print()
def findingFittingError(truth_h5, measured_h5, pixel_size = None, max_distance = None):
"""
truth_h5 - A saH5Py.SAH5Py object with the ground truth localizations.
measured_h5 - A saH5Py.SAH5Py object with the found localizations.
pixel_size - The camera pixel size in nanometers. If not specified then the value
in the measured HDF5 file will be used.
max_distance - If not none, found peaks that are greater than this distance from
a truth peak will be ignored. Units are nanometers.
"""
if (measured_h5.getNLocalizations() == 0):
return [None, None, None]
md_in_pixels = None
md_sqr = None
if max_distance is not None:
md_in_pixels = max_distance/pixel_size
md_sqr = max_distance * max_distance
if pixel_size is None:
pixel_size = measured_h5.getPixelSize()
all_dx = []
all_dy = []
all_dz = []
for i in range(truth_h5.getMovieLength()):
t_locs = truth_h5.getLocalizationsInFrame(i)
m_locs = measured_h5.getLocalizationsInFrame(i)
if not bool(t_locs) or not bool(m_locs):
continue
p_index = iaUtilsC.peakToPeakDistAndIndex(m_locs['x'], m_locs['y'],
t_locs['x'], t_locs['y'],
max_distance = md_in_pixels)[1]
for i in range(m_locs['x'].size):
if(p_index[i] < 0):
continue
dx = pixel_size * (m_locs['x'][i] - t_locs['x'][p_index[i]])
dy = pixel_size * (m_locs['y'][i] - t_locs['y'][p_index[i]])
if 'z' in m_locs:
dz = 1000.0 * (m_locs['z'][i] - t_locs['z'][p_index[i]])
else:
dz = 0.0
if md_sqr is not None:
if ((dx*dx + dy*dy + dz*dz) < md_sqr):
all_dx.append(dx)
all_dy.append(dy)
all_dz.append(dz)
else:
all_dx.append(dx)
all_dy.append(dy)
all_dz.append(dz)
return [numpy.array(all_dx), numpy.array(all_dy), numpy.array(all_dz)]
def collate():
dirs = sorted(glob.glob("test*"))
if (len(dirs) == 0):
print("No test directories found.")
exit()
# Load reference localizations.
ref = saH5Py.loadLocalizations("sim_input_c1_grid_list.hdf5",
fields=["color", "x", "y"])
for a_dir in dirs:
# Check color correspondence.
#
# Note: Currently only works for gridded localizations.
#
exp = saH5Py.loadTracks(a_dir + "/test.hdf5",
fields=["x", "y", "km_color"])
# Identify corresponding peaks with 2 pixel maximum radius.
p_index = iaUtilsC.peakToPeakDistAndIndex(exp["x"],
exp["y"],
ref["x"],
ref["y"],
max_distance=2.0)[1]
# Create array for checking category correspondence.
c_size = numpy.count_nonzero(p_index > -1)
categories = numpy.zeros((c_size, 2), dtype=numpy.int32)
i = 0
for j in range(p_index.size):
if (p_index[j] < 0):
continue
categories[i, 0] = int(ref["color"][p_index[j]])
categories[i, 1] = int(exp["km_color"][j])
i += 1
# Measure fraction of experimental localizations assigned correctly. This
# is complicated by the k-mean category being arbitrary.
for i in range(8):
ref_mask = (categories[:, 0] == i)
if (numpy.count_nonzero(ref_mask) > 0):
exp_cat = int(numpy.median(categories[ref_mask, 1]))
exp_mask = (categories[:, 1] == exp_cat)
total = numpy.count_nonzero(exp_mask)
matched = numpy.count_nonzero(
numpy.logical_and(ref_mask, exp_mask))
mismatched = total - matched
print(
"Color {0:0d}, matching fraction is {1:.2f}, unmatched fraction is {2:.2f}, total {3:0d}"
.format(i, matched / total, mismatched / total, total))
def collate():
dirs = sorted(glob.glob("test*"))
if(len(dirs) == 0):
print("No test directories found.")
exit()
# Load reference localizations.
ref = saH5Py.loadLocalizations("sim_input_c1_grid_list.hdf5", fields = ["color", "x", "y"])
for a_dir in dirs:
# Check color correspondence.
#
# Note: Currently only works for gridded localizations.
#
exp = saH5Py.loadTracks(a_dir + "/test.hdf5", fields = ["x", "y", "km_color"])
# Identify corresponding peaks with 2 pixel maximum radius.
p_index = iaUtilsC.peakToPeakDistAndIndex(exp["x"], exp["y"], ref["x"], ref["y"],
max_distance = 2.0)[1]
# Create array for checking category correspondence.
c_size = numpy.count_nonzero(p_index > -1)
categories = numpy.zeros((c_size, 2), dtype = numpy.int32)
i = 0
for j in range(p_index.size):
if (p_index[j] < 0):
continue
categories[i,0] = int(ref["color"][p_index[j]])
categories[i,1] = int(exp["km_color"][j])
i += 1
# Measure fraction of experimental localizations assigned correctly. This
# is complicated by the k-mean category being arbitrary.
for i in range(8):
ref_mask = (categories[:,0] == i)
if(numpy.count_nonzero(ref_mask) > 0):
exp_cat = int(numpy.median(categories[ref_mask,1]))
exp_mask = (categories[:,1] == exp_cat)
total = numpy.count_nonzero(exp_mask)
matched = numpy.count_nonzero(numpy.logical_and(ref_mask, exp_mask))
mismatched = total - matched
print("Color {0:0d}, matching fraction is {1:.2f}, unmatched fraction is {2:.2f}, total {3:0d}".format(i, matched/total, mismatched/total, total))
def collateDAO(dirs, settings, calc_width_error = True):
"""
Results collations for 3D-DAOSTORM and sCMOS analysis.
"""
all_dx = []
all_dy = []
all_wx = []
all_wy = []
noise = 0
noise_total = 0
recall = 0
recall_total = 0
total_locs = 0
total_time = 0.0
for a_dir in dirs:
print("Processing", a_dir)
# Load timing information.
with open(a_dir + "/timing.txt") as fp:
total_time += float(fp.readline())
# Load localizations.
truth_h5 = saH5Py.SAH5Py(a_dir + "/test_ref.hdf5")
measured_h5 = saH5Py.SAH5Py(a_dir + "/test.hdf5")
total_locs += measured_h5.getNLocalizations()
# Calculate fractional recall.
[partial, total] = rfrac.recallFraction(truth_h5, measured_h5, settings.tolerance)
recall += partial
recall_total += total
# Calculate noise fraction.
[partial, total] = rfrac.noiseFraction(truth_h5, measured_h5, settings.tolerance)
noise += partial
noise_total += total
# Calculate error in fitting width.
if calc_width_error:
for i in range(truth_h5.getMovieLength()):
t_locs = truth_h5.getLocalizationsInFrame(i)
m_locs = measured_h5.getLocalizationsInFrame(i)
# Widths for truth localizations.
t_wx = t_locs["xsigma"]
t_wy = t_locs["ysigma"]
# Widths for found localizations.
if ("xsigma" in m_locs):
m_wx = m_locs["xsigma"]
else:
m_wx = 1.5*numpy.ones(m_locs['x'].size)
if ("ysigma" in m_locs):
m_wy = m_locs["ysigma"]
else:
m_wy = m_wx.copy()
p_index = iaUtilsC.peakToPeakDistAndIndex(m_locs['x'], m_locs['y'],
t_locs['x'], t_locs['y'],
max_distance = settings.tolerance)[1]
p_size = numpy.count_nonzero(p_index > -1)
d_wx = numpy.zeros(p_size)
d_wy = numpy.zeros(p_size)
k = 0
for j in range(m_locs["x"].size):
if(p_index[j] < 0):
continue
d_wx[k] = m_wx[j] - t_wx[p_index[j]]
d_wy[k] = m_wy[j] - t_wy[p_index[j]]
k += 1
all_wx.append(d_wx)
all_wy.append(d_wy)
# Calculate fitting error in XY.
max_distance = None
if True:
max_distance = 2.0 * settings.pixel_size
print("Using max_distance", max_distance, "nm for error calcuations.")
[dx, dy, dz] = ffe.findingFittingError(truth_h5,
measured_h5,
pixel_size = settings.pixel_size,
max_distance = max_distance)
if dx.size != 0:
all_dx.append([numpy.std(dx), math.sqrt(numpy.mean(dx*dx))])
all_dy.append([numpy.std(dy), math.sqrt(numpy.mean(dy*dy))])
else:
all_dx.append([0,0])
all_dy.append([0,0])
truth_h5.close()
measured_h5.close()
print()
print("Analysis Summary:")
print("Processed {0:0d} localizations in {1:.2f} seconds, {2:.2f}/sec".format(total_locs, total_time, float(total_locs)/float(total_time)))
print("Recall {0:.5f}".format(float(recall)/float(recall_total)))
print("Noise {0:.5f}".format(float(noise)/float(noise_total)))
print("XY Error Standard Deviation (nm):")
for i, a_dir in enumerate(dirs):
print(a_dir + "\t{0:.2f}\t{1:.2f}".format(all_dx[i][0], all_dy[i][0]))
print("")
print("XY RMSE (nm):")
for i, a_dir in enumerate(dirs):
print(a_dir + "\t{0:.2f}\t{1:.2f}".format(all_dx[i][1], all_dy[i][1]))
if calc_width_error:
print("")
print("XY Width Error, Mean difference with truth, Standard deviation (pixels):")
for i, a_dir in enumerate(dirs):
print(a_dir + "\t{0:.3f}\t{1:.3f}\t{2:.3f}\t{3:.3f}".format(numpy.mean(all_wx[i]),
numpy.std(all_wx[i]),
numpy.mean(all_wy[i]),
numpy.std(all_wy[i])))
def findingFittingError(truth_h5,
measured_h5,
pixel_size=None,
max_distance=None):
"""
truth_h5 - A saH5Py.SAH5Py object with the ground truth localizations.
measured_h5 - A saH5Py.SAH5Py object with the found localizations.
pixel_size - The camera pixel size in nanometers. If not specified then the value
in the measured HDF5 file will be used.
max_distance - If not none, found peaks that are greater than this distance from
a truth peak will be ignored. Units are nanometers.
"""
if (measured_h5.getNLocalizations() == 0):
return [None, None, None]
md_in_pixels = None
md_sqr = None
if max_distance is not None:
md_in_pixels = max_distance / pixel_size
md_sqr = max_distance * max_distance
if pixel_size is None:
pixel_size = measured_h5.getPixelSize()
all_dx = []
all_dy = []
all_dz = []
for i in range(truth_h5.getMovieLength()):
t_locs = truth_h5.getLocalizationsInFrame(i)
m_locs = measured_h5.getLocalizationsInFrame(i)
if not bool(t_locs) or not bool(m_locs):
continue
p_index = iaUtilsC.peakToPeakDistAndIndex(m_locs['x'],
m_locs['y'],
t_locs['x'],
t_locs['y'],
max_distance=md_in_pixels)[1]
for i in range(m_locs['x'].size):
if (p_index[i] < 0):
continue
dx = pixel_size * (m_locs['x'][i] - t_locs['x'][p_index[i]])
dy = pixel_size * (m_locs['y'][i] - t_locs['y'][p_index[i]])
if 'z' in m_locs:
dz = 1000.0 * (m_locs['z'][i] - t_locs['z'][p_index[i]])
else:
dz = 0.0
if md_sqr is not None:
if ((dx * dx + dy * dy + dz * dz) < md_sqr):
all_dx.append(dx)
all_dy.append(dy)
all_dz.append(dz)
else:
all_dx.append(dx)
all_dy.append(dy)
all_dz.append(dz)
return [numpy.array(all_dx), numpy.array(all_dy), numpy.array(all_dz)]
