# !/usr/bin/python import os from neuralimg.image import preproc as pr from neuralimg.crag.merge_mc import * """ Generates superpixels given the membrane probabilities """ mask = 7 sigma = 0.25 sps_folder = 'data/crag/sps' mems_folder = 'data/crag/mem' mems_norm = 'data/crag/mem_norm' # Extract images proc = pr.DatasetProc(mems_folder) proc.read() proc.segment(mask, sigma) proc.save_data(sps_folder) # Generate normalized membranes [0,1] proc = pr.DatasetProc(mems_folder) proc.read() proc.normalize() proc.save_data(mems_norm) # MC Extractor mc_path = os.path.join(sps_folder, 'histories_mc') MCTreeExtractor(sps_folder, mems_norm).extract(mc_path)
def get_threshold(mask,
sigma,
mems,
truth,
ted_shift,
split_bg=False,
mweight=10,
merge_values=15,
stop=False,
ted_workers=1,
tmp_path=None):
""" Gets the best threshold in the merge trees computed on the segmentation
given by the input parameters
Args:
masks: List of masks to test
sigmas: List of sigmas to test
membranes: Membrane predictions
truth: Folder containing groundtruth images
ted_shift: minimum shift in the neuron boundaries for considering errors
split_bg: Whether to split between errors and background errors
mweight: weight to apply to merges. To be between 1 and 20.
merge_values: Number of cuts to test for the merge tree histories
stop: Whether to stop when first minimum found for a given threshold value.
workers: Number of workers to use. Default: 3
ted_workers: Number of workers to use for each image evaluation. Should be kept
low since the total number of threads will be of the order of
workers * ted_workers
tmp_path: Folder to use as temporary path for intermediate files. Set to None
to use OS default.
Return:
best setting, all settings, sigma and mask
"""
logging.info('Evaluating sigma %f and mask %d' % (sigma, mask))
# Prepare temporary folder
if tmp_path is not None:
create_dir(tmp_path)
outp = os.path.join(tmp_path,
'_'.join(['s', str(sigma), 'm',
str(mask)]))
create_dir(outp)
else:
outp = tempfile.mkdtemp()
sps = os.path.join(outp, 'sps')
hists = os.path.join(outp, 'hists')
# Save superpixels
proc = pr.DatasetProc(mems)
proc.read()
proc.segment(mask, sigma)
proc.save_data(sps)
# Create merge trees
mt.MCTreeExtractor(sps, mems).extract(hists)
# Get optimal threshold
best, all_stats = search_threshold(sps,
truth,
hists,
ted_shift,
mweight,
merge_values,
split_bg,
out_stats=None,
stop=stop,
ted_workers=ted_workers)
# Clean temporary folder
shutil.rmtree(outp)
return best, all_stats, sigma, mask
def generate_superpixels(mask, sigma, mems, sps):
""" Generates the superpixels in the given folder according to parameters """
proc = pr.DatasetProc(mems)
proc.read()
proc.segment(mask, sigma)
proc.save_data(sps)
# !/usr/bin/python
from neuralimg.image import preproc as pr
import matplotlib.pyplot as plt
""" Generates superpixels given the membrane probabilities """
proc = pr.DatasetProc('/DataDisk/morad/cremi/superpixels/sampleA/mask3_sigma1_ted25')
proc.read()
sizes, mean, std = proc.compute_supervoxel_stats()
print('Mean: %f, Standard deviation: %f' % (mean, std))
plt.hist(sizes, 50, normed=1, histtype='stepfilled')
plt.show()
# ! /usr/bin/python
from neuralimg.image import preproc as pr
""" Normalizes the membrane probabilities so they are in [0,1] interval """
proc = pr.DatasetProc('data/crag/mem')
proc.read()
proc.normalize()
proc.save_data('data/crag/mem_norm')
# !/usr/bin/python
from neuralimg.image import preproc
""" Preprocesses groundtruth so it joins small supervoxels with neighboring ones
and gives separate ids to unconnected regions even if they belong to the same neuron
(for ease of 2D evaluation) """
# From folder
p = preproc.DatasetProc('data/crag/gt')
p.read()
p.join_small()
p.save_data('data/crag/gt_large')
p.split_labels()
p.save_data('data/crag/gt_large_unconnected')